June 2026 Summaries
93 posts from Hugging Face
Filter
Month:
Year:
Post Summaries
Back to Blog
Every Eval Ever (EEE) and Hugging Face Community Evals have become intercompatible to enhance the reporting and interpretation of AI evaluation results, aiming to address the inconsistencies and scattered nature of current evaluation practices. Launched in February 2026, these initiatives allow cross-posting and provide a unified standardized metadata store, improving trust and comprehension for users, researchers, and policymakers. EEE offers a JSON schema to standardize evaluation reporting, capturing details like generation settings and metrics, while Hugging Face Community Evals decentralizes the reporting of benchmark scores, compiling them into a comprehensive leaderboard system. The collaboration enables contributors to submit evaluation data that appear both on Hugging Face model pages and within the EEE records, ensuring results are accessible and interpretable through integrated Eval Cards. This system facilitates efficient evaluation result management, mitigating duplication and enhancing transparency by linking each score to its detailed source record, with the converter tool simplifying the process of integrating data into both platforms.
Jun 30, 2026
1,434 words in the original blog post.
Specialization is increasingly recognized as a crucial principle in the development of effective AI systems, as highlighted by the 2026 work of Goldfeder, Wyder, LeCun, and Shwartz-Ziv. This perspective is supported by optimization theory, evolutionary biology, competitive markets, and machine learning, all of which suggest that systems achieve superior performance by focusing narrowly on specific tasks rather than attempting to be general-purpose. The "No Free Lunch" theorem mathematically supports this by showing that no single algorithm performs best across all problems, implying that resources should be concentrated on a finite set of tasks for optimal results. This idea is mirrored in biological evolution, where specialization is favored due to limited resources, and in competitive markets, where organizations that target specific niches outperform generalists. Machine learning experiences similar patterns, as systems trained on multiple tasks can suffer from negative transfer unless tasks are inherently cooperative. The article further argues that scaling in AI will not eliminate the need for specialization, as it pertains to directing resources effectively rather than encoding domain-specific knowledge, and concludes that specialization is an emergent property of constrained systems striving for performance.
Jun 30, 2026
2,264 words in the original blog post.
ScarfBench is an open benchmark designed to evaluate AI agents' abilities in migrating enterprise Java applications across frameworks like Spring, Jakarta EE, and Quarkus, addressing a significant yet challenging area of software engineering. Unlike traditional benchmarks that focus on code generation, ScarfBench emphasizes not just code translation but also the preservation of application behavior, successful deployment, and dependency management, which are crucial for real-world applications. Despite advances in coding agents, the benchmark reveals that framework migration remains difficult, with current AI agents achieving low success rates in preserving application behavior. This difficulty is compounded by complex dependencies and environmental challenges, such as configuration and tool inconsistencies, highlighting the need for reliable validation and architectural reasoning. ScarfBench offers a comprehensive dataset, evaluation infrastructure, and a public leaderboard, serving as a valuable resource for researchers and practitioners aiming to improve AI-assisted application modernization while encouraging contributions of new migration scenarios and techniques.
Jun 30, 2026
1,067 words in the original blog post.
Heretic Grimoire is an application designed to archive, organize, and provide access to Heretic-generated models' reproduce.json files, making it possible to recreate language models even if the original sources are removed. Operating on a twice-daily schedule, it collects and indexes new records from Hugging Face, ensuring that previously stored data remains accessible. The app allows users to search metadata, download archives, and maintain a local or cloud-based backup, which is critical given the potential risks of centralized model hosting and legal challenges like the one faced from Meta. Heretic models, known for their decensored nature, can be reproduced using minimal data, with exact replication dependent on matching system configurations. Since version 1.4.0, Heretic has automated the reproduction process, allowing users to easily restore models with the heretic --reproduce command, enhancing the utility and reliability of this archiving system.
Jun 30, 2026
1,360 words in the original blog post.
Chitos, developed by VIDRAFT, is an autonomous security AI designed to bridge the gap between vulnerability detection and proof by actively demonstrating exploits. Unlike static analysis tools, Chitos uses a three-phase pipeline: it starts with static analysis to identify potential threats, then engages in autonomous research using web searches to verify these threats, and finally performs live attacks on authorized targets to provide concrete evidence of vulnerabilities. This approach mitigates the common issue of false positives by ensuring each finding is substantiated with proof, rather than mere suspicion. Chitos operates on VIDRAFT's Darwin-398B-JGOS model, known for its multi-hop logical chaining capabilities, and offers features such as dynamic reasoning and live process streaming for vulnerabilities like SQL injection and cross-site scripting. The tool emphasizes responsible use, requiring users to have ownership or explicit authorization to test target systems, and supports safe practice environments for testing without legal concerns.
Jun 29, 2026
1,660 words in the original blog post.
The Multimodal Universe (MMU) consolidates over 80TB of astronomical data from more than 30 surveys into a user-friendly format, facilitating crossmatching, which links observations of the same object across different surveys. Previously, this process required significant local storage, but a recent conversion to the parquet-based HATS format, accessible via the LSDB and Hugging Face ecosystems, allows astronomers to perform crossmatches on laptops with just 4GB of RAM. This advancement democratizes access to powerful astronomical data processing, enabling researchers to conduct complex analyses without needing high-end hardware. Crossmatching plays a crucial role in identifying unique astronomical phenomena and testing hypotheses like the Platonic Representation Hypothesis, which explores the convergence of neural networks on a unified model of reality. The transformation of MMU into HATS format, supported by the LINCC Frameworks, allows efficient streaming and spatial operations, thereby enhancing the usability of astronomical data and fostering broader participation in scientific discovery.
Jun 29, 2026
2,494 words in the original blog post.
DiScoFormer, a novel transformer model, is introduced as a solution for estimating both the density and score of data distributions in a single forward pass without retraining, overcoming limitations in traditional methods like kernel density estimation (KDE) and neural score-matching models. By leveraging cross-attention and a shared backbone with separate output heads for density and score, DiScoFormer can evaluate these metrics at any point, maintaining accuracy even in high-dimensional spaces where KDE struggles. The model is trained using Gaussian Mixture Models, which provide exact targets for supervision due to their universal density approximation capabilities. DiScoFormer significantly outperforms KDE in both density and score estimation, particularly in high dimensions, and demonstrates adaptability to out-of-distribution inputs without requiring ground-truth data. Its promise lies in its ability to serve as a plug-in estimator that remains accurate across various applications, such as generative modeling and Bayesian inference, reducing the need for retraining across different problems.
Jun 29, 2026
894 words in the original blog post.
Gemma-4 31B, a dense Transformer model developed by Google DeepMind, was evaluated on the vLLM serving engine using an RTX 6000 PRO GPU to benchmark its performance under varying concurrency levels from 12 to 24. The model, optimized for reasoning, coding, and multimodal understanding, was tested with a 4K-token context window to align with ShareGPT dataset requirements. Results showcased impressive throughput peaking at 1.17k tokens per second, with a median time to first token (TTFT) of approximately 0.7 seconds, although tail latency presented a challenge under heavy load, reaching up to 19 seconds for the p99 metric. Despite this, the server maintained low queue depths, indicating efficient handling of requests even at maximum concurrency, with long end-to-end latencies attributed primarily to the generation of lengthy outputs rather than server inefficiencies. The benchmarking was conducted using HexGrid.cloud, highlighting the platform's capability to deploy open models on dedicated GPUs effectively.
Jun 29, 2026
786 words in the original blog post.
VLX-Go is a compact vision-language waypoint planner designed to enhance embodied navigation by predicting short-horizon local waypoints based on recent visual frames, current observations, and natural-language instructions. It addresses the challenge of transforming multimodal inputs into actionable navigation targets for robots, focusing on local motion rather than global route planning. This lightweight model operates in a closed-loop system, allowing for dynamic updates and corrections based on real-time observations, making it suitable for tasks like target following and obstacle avoidance. By separating high-level waypoint prediction from low-level control, VLX-Go provides a practical interface for integrating planning with safety checks and simulator feedback, facilitating easier deployment and evaluation in real-world robotic systems. The model is trained using a combination of offline trajectory data and online simulator feedback to enhance its robustness against obstacles and drift. VLX-Go achieves strong performance metrics, notably in navigation success and target tracking, while maintaining a deployable structure for closed-loop systems.
Jun 28, 2026
1,138 words in the original blog post.
VLX-Flow is a novel model designed for real-time video understanding, addressing the limitations of traditional video models that wait for user queries before processing. Unlike offline workflows which require reprocessing entire video histories, VLX-Flow continuously processes video streams in chronological chunks, updating its internal memory incrementally. This allows it to answer questions based on a maintained state without rewatching the video, making it more efficient for live environments. The model uses a two-layer memory system, with a visual cache for short-term details and semantic memory for higher-level context, ensuring stable latency and smoother memory growth. This approach supports real-time video question answering and event-triggered interactions, making it suitable for edge devices where bandwidth, latency, and privacy are concerns. VLX-Flow transforms video understanding into a continuously running perception module, ideal for devices that need to process video as a live, ongoing context.
Jun 27, 2026
1,223 words in the original blog post.
VLX-Seek is an innovative model designed to enhance fine-grained perception in multimodal large models (VLMs) for real-world applications like cameras, drones, and robots, shifting focus from generating coordinate-based localization to region reference. Unlike traditional VLMs that excel in semantic understanding yet struggle with precise localization, VLX-Seek employs a novel approach by using region tokens, which allows the model to refer to specific parts of an image as language entities. This method enhances the model’s ability to perform tasks such as object detection, open-vocabulary localization, and complex referring expression comprehension by turning localization into a language-conditioned retrieval among candidate visual regions. This approach not only improves inference efficiency and accuracy but also reduces the computational demands, making it especially suitable for on-device applications where resources are limited. As a result, VLX-Seek empowers embodied systems to execute actions based on stable and accurate spatial anchors, thus bridging the gap between image understanding and actionable perception.
Jun 27, 2026
3,375 words in the original blog post.
DukaanBench is an innovative AI benchmark that challenges language models to operate a simulated Indian kirana store for 30 days, assessing their ability to manage inventory, cash, customer trust, and marketing strategies. Each day, the model receives a comprehensive state of the shop and must return an executable JSON action to guide store operations, with the backend simulating customer interactions and updating variables like trust and inventory. The benchmark aims to evaluate not just profit-making capabilities but also the model's ability to maintain operational stability and customer relationships, with metrics including service rate, trust, and marketing effectiveness. The initial findings highlight the importance of aligning action with rationale, managing trust, and ensuring inventory awareness in marketing efforts. Part 1 introduces the environment and evaluation loop, while Part 2 will explore training a smaller, more specialized model to improve on these tasks, offering potential as a practical tool for shopkeepers rather than replacing them.
Jun 27, 2026
3,871 words in the original blog post.
Quentin Gallouédec provides a guide on deploying a private, OpenAI-compatible large language model (LLM) endpoint on Hugging Face's infrastructure using a single command, eliminating the need for manual server provisioning and Kubernetes management, and offering a pay-per-second billing model. The setup involves using the official vllm/vllm-openai image, requesting a GPU, and exposing the model's port through Hugging Face's public jobs proxy for easy access from any location via an API token. It caters to various use cases such as tests, evaluations, and batch generation, and details how to scale the command for larger models, use curl or Python for queries, and secure access with an HF token. The post also explains additional functionalities like integrating with Gradio for a UI chat interface, SSH access for debugging, and utilizing the endpoint as a coding-agent backend with Pi. It provides a comparison between Hugging Face Jobs and Inference Endpoints, recommending the former for flexibility and experiments, and the latter for production-ready, long-term services with enhanced access control and operational features.
Jun 26, 2026
1,611 words in the original blog post.
VLX-Flow represents a significant advancement in video understanding by enabling continuous, real-time multimodal interaction, addressing the limitations of traditional offline models which process videos only after a query is made. This system processes video streams as sequences of streaming chunks, updating its internal memory incrementally to maintain an evolving visual state, thus allowing it to answer questions from the accumulated context without reprocessing the entire video history. By using Linear Attention and a two-layer memory approach, VLX-Flow ensures stable latency and efficient memory usage, preserving both short-term visual details and long-term semantic context. This supports real-time video question answering and event-triggered interactions, making it particularly valuable for on-device and edge scenarios, where bandwidth, latency, and privacy are concerns. Ultimately, VLX-Flow transforms video understanding into a continuously running perception module, aligning more closely with the persistent observational nature of real-world devices like cameras and robots.
Jun 26, 2026
1,194 words in the original blog post.
OlmoLogic is a novel model designed to enhance reasoning capabilities by integrating Inductive Logic Programming (ILP) into the Olmo-3 Reinforcement Learning via Verifiers (RLVR) framework, focusing on logical reasoning, which is often neglected in favor of math and code. The model was trained intensively on 56 H100 GPUs over six days, aiming to improve its logical reasoning skills through tasks from the Scalable Logical Reasoning (SLR) suite, which includes a diverse set of 19,000 tasks that vary in complexity. OlmoLogic achieved significant improvements, tripling the accuracy on the SLR-Bench and showing gains across various logic benchmarks while maintaining performance in math, code, and instruction-following tasks. The training incorporated a Prolog interpreter to execute the logic programs proposed by the model, providing direct feedback used as RLVR rewards, and introduced a reward structure that emphasizes rule simplicity and correctness. The development also included Olmo 3.1 7B Think, a variant trained without SLR tasks for comparison, highlighting the impact of SLR on logical reasoning. Overall, OlmoLogic represents a significant step in integrating logical reasoning into AI models, providing a robust framework for reasoning tasks without altering the underlying training infrastructure.
Jun 26, 2026
2,702 words in the original blog post.
The SportsBERT Small series comprises domain-specialized language models, optimized to perform effectively in sports-related tasks with only 22.7 million parameters, showcasing a performance that rivals much larger models. These models, trained using masked language modeling on sports-labeled Wikipedia articles, demonstrate that focusing on a specific domain can lead to efficient models that require fewer parameters than those designed for broader applications. The series includes variations like the SportsBERT Small Base and small embeddings models, with the latter being fine-tuned for generating vector embeddings through distillation from larger models. Evaluation results highlight that despite its compact size, SportsBERT Small Embeddings outperforms similarly sized models and is competitive with much larger ones, proving advantageous for CPU-only environments where computational resources and disk space are limited. This development underscores the potential of small, specialized models in achieving high efficiency and accuracy within specific domains, facilitated by NeuML, the company providing AI consulting services and developing txtai applications.
Jun 26, 2026
1,126 words in the original blog post.
The study examines the prediction capabilities of hybrid language models compared to standard transformers, focusing on token-level differences. Conducted with the Olmo Hybrid and Olmo 3 models, the research reveals that hybrids excel in predicting meaning-bearing tokens such as nouns, verbs, and adjectives, as well as tokens requiring contextual understanding, like pronouns. However, their advantage diminishes on repeated tokens, where transformers excel due to their attention mechanism's ability to recall specific earlier tokens. The hybrid model's strength lies in its recurrent layers' ability to track state changes, though it struggles with precise recall. The findings suggest that evaluating models based on specific token types rather than overall loss provides a clearer picture of architectural strengths and weaknesses, particularly highlighting the hybrid's proficiency with open-class tokens. The research encourages deeper exploration of token-specific losses to enhance model development, with the aim of refining hybrid architectures by understanding each component's unique capabilities.
Jun 25, 2026
1,364 words in the original blog post.
Moon Bot is an innovative engineering assistant developed by HuggingFace to streamline workflows within Slack by integrating with various tools like Elasticsearch and MongoDB without requiring users to switch contexts. By utilizing HuggingFace's infrastructure, Moon Bot allows seamless access to metrics, codebases, and other resources through simple Slack messages, thus eliminating the need for multiple authentication processes and interfaces. It operates using the Pi coding agent SDK, running in a Kubernetes pod, and utilizes HuggingFace Buckets for persistent memory storage, allowing it to resume conversations even after restarts. The bot is designed with a focus on security, using tiered access and sandboxed execution to protect sensitive data, and it facilitates actions like opening GitHub pull requests without exposing write access. Moon Bot's architecture supports pluggable skills via Markdown files that guide its interactions with different tools, ensuring flexibility and easy updates. It also runs scheduled tasks to generate reports and monitor deployments, showcasing a replicable model for integrating AI-driven agents in organizational workflows.
Jun 24, 2026
2,003 words in the original blog post.
Interhuman's Inter-1 model, designed to interpret human communication from videos, displayed a peculiar behavior of fabricating quotes like "Yeah, Friday at five" when audio was missing. This anomaly, attributed to a fallback mechanism where the model draws from its training to fill gaps, was traced back to a specific example in the system prompt rather than the training data itself. The investigation revealed that the model's tendency to invent speech in silence was a result of learned priors, demonstrating a broader "Clever Hans" effect, where models rely on prior expectations rather than actual data to make judgments. Efforts to address this issue have focused on modifying prompts and understanding the model's learned behaviors, highlighting the challenge of making omni-modal models robust when data from a modality is absent. The research team continues to explore solutions to ensure that missing modalities are treated as such, rather than prompting fabricated responses.
Jun 24, 2026
2,371 words in the original blog post.
NVIDIA NeMo AutoModel, an open library within the NVIDIA NeMo framework, significantly enhances the efficiency of fine-tuning Mixture-of-Experts (MoE) models by integrating seamlessly with HuggingFace Transformers v5. It introduces Expert Parallelism, DeepEP fused all-to-all dispatch, and TransformerEngine kernels, resulting in a 3.4-3.7x increase in training throughput and a reduction of GPU memory usage by 29-32% compared to native Transformers v5. The integration is designed to maintain API compatibility with HuggingFace, requiring only a single import line change to leverage these improvements. This setup allows for scalable training across multiple GPUs, making it feasible to fine-tune large models like the 550B-parameter Nemotron 3 Ultra across 16 nodes. NeMo AutoModel's optimizations include sharding expert weights across GPUs and fusing communication with computation to enhance speed and efficiency, all while maintaining compatibility with standard HF-format checkpoints for easy deployment on various inference frameworks.
Jun 24, 2026
2,234 words in the original blog post.
The FFASR Leaderboard, launched by Treble Technologies and Hugging Face, is the first open benchmark designed to evaluate automatic speech recognition (ASR) models under realistic far-field acoustic conditions. This community-driven initiative addresses the significant performance gap between standard near-field evaluations and real-world scenarios involving complex acoustics such as reverberation and background noise. By simulating diverse environments across 14 different room types, the benchmark provides a standardized framework for assessing models on far-field performance, emphasizing the importance of acoustic robustness in voice interfaces that operate in challenging environments. The leaderboard evaluates models across various conditions, including different signal-to-noise ratios (SNRs) and moving-source scenarios, allowing researchers to understand the trade-offs between accuracy and speed. The initiative aims to encourage the development of ASR models that are resilient to real-world acoustic challenges and invites the community to contribute models and insights to further refine the benchmark.
Jun 24, 2026
1,647 words in the original blog post.
Kog, a Paris-based AI infrastructure startup, has released Laneformer 2B, a 2.3 billion parameter coding model optimized for high-speed decoding, on the Hugging Face Hub. Unlike traditional approaches that prioritize benchmark quality, Kog focused on maximizing inference speed from the outset, designing the model and its architecture to integrate seamlessly with their Kog Inference Engine. This latency-first approach led to the development of Delayed Tensor Parallelism (DTP), which delays inter-GPU communication costs, enhancing decoding speed without compromising model quality. Laneformer 2B, trained with a mixture of open-source data, demonstrates competitive coding capabilities, achieving high scores on benchmarks like HumanEval+ and MBPP+. Kog's open-source release includes the model weights, architecture, and documentation, aiming to encourage community involvement and innovation in latency-oriented model design. The model's training leveraged efficient European infrastructure and high-performance GPUs, ensuring a robust and repeatable process.
Jun 24, 2026
3,042 words in the original blog post.
Huggingface_hub, the Python client integral to the Hugging Face ecosystem, has shifted from a release cycle of every 4 to 6 weeks to weekly releases, facilitated by a streamlined, largely automated process using open-source tools. This new workflow leverages GitHub Actions and incorporates AI to draft release notes, which are then reviewed by humans to ensure accuracy and tone, maintaining a "trust but verify" approach. The process includes automation of mechanical tasks like version bumping and publishing while preserving human oversight for tasks requiring judgment, such as crafting release notes and internal announcements. The open and reusable nature of this system allows other maintainers to adapt it for their projects, emphasizing a collaborative and transparent development environment. The integration of deterministic checks ensures the reliability of AI-generated content, significantly reducing the workload on human contributors while maintaining high-quality outputs.
Jun 23, 2026
2,343 words in the original blog post.
Transformers.js, a tool for integrating transformer models into web applications, faces challenges with cache redundancy when deploying AI models across different origins. This issue arises because browsers isolate caches by origin to enhance privacy and security, resulting in duplicate downloads and storage for identical models and WebAssembly (Wasm) runtime resources. The proposed Cross-Origin Storage (COS) API aims to address this by allowing apps to share resources across origins using cryptographic hashes rather than URLs, thus preventing redundant downloads. COS enables developers to specify which origins can access stored resources, maintaining privacy controls while optimizing resource sharing. Although not yet natively implemented in browsers, developers can experiment with the COS API through a browser extension. This approach has been piloted in Transformers.js, showing significant reductions in duplicate data transfers, and the Chrome team is considering native implementation of the API.
Jun 23, 2026
2,915 words in the original blog post.
CUGA (Configurable Generalist Agent) is an open-source agent harness developed by IBM to streamline the creation and deployment of agentic applications by handling complex tasks such as planning, execution, and state management, allowing developers to focus on specifying tools and prompts. By providing a pre-assembled infrastructure, CUGA reduces the need for extensive setup traditionally required in building agentic apps, offering a library of two-dozen single-file applications as examples. These applications, built using FastAPI, demonstrate various use cases from movie recommendations to cloud architecture advising and highlight the reusability and adaptability of the CUGA framework. CUGA's design emphasizes governance and scalability, featuring a policy system and multi-agent delegation to ensure secure and efficient operations, with options for deploying agents in secure environments like IBM's Sovereign Core. The system's flexibility allows for easy swapping of underlying models and tools, making it adaptable to different operational contexts while maintaining predictability and safety through built-in governance features.
Jun 23, 2026
3,392 words in the original blog post.
ThousandWorlds is a comprehensive benchmark designed to aid the machine learning and exoplanet communities in studying alien climates by providing a dataset of 1,760 simulations from five global climate models (GCMs), supplemented with additional bespoke runs. These simulations cover a range of planetary conditions, from icy to sauna-like environments, and are used to perform parameter-to-field regression, predicting a planet's 3D climate based on eight parameters. The ThousandWorlds dataset addresses the challenge of limited data, parameter-to-field prediction, and varying fidelity across multiple simulators, offering an opportunity for the ML community to explore and benchmark methods in an area of scientific research that lacks dominant deep learning solutions. The dataset's design considers spherical geometry and structured missingness, providing a platform to test and improve models, such as Gaussian processes and deep learning approaches, for more accurate climate predictions on exoplanets.
Jun 23, 2026
899 words in the original blog post.
PP-OCRv6 is the latest iteration of PaddleOCR's universal OCR models, designed to enhance text detection and recognition across various real-world scenarios, from documents to industrial labels. It introduces three model tiers—tiny, small, and medium—ranging from 1.5M to 34.5M parameters, supporting up to 50 languages, including Chinese, English, Japanese, and Latin-script languages. The update brings architectural, training, and data improvements, achieving better detection and recognition accuracy compared to its predecessor, PP-OCRv5_server. Key features include the PPLCNetV4 backbone for consistency across all tiers and the use of RepLKFPN and EncoderWithLightSVTR for efficient text detection and recognition. The models can be integrated with PaddlePaddle, Transformers, or ONNX Runtime backends, offering flexibility for different deployment environments. PP-OCRv6 is available for evaluation and integration through an online demo, model collection, and various inference backends, making it suitable for diverse OCR needs in multilingual and complex text scenarios.
Jun 22, 2026
1,089 words in the original blog post.
In June 2026, the importance of owning and running AI models locally was emphasized following the removal of Anthropic's Claude Fable 5, highlighting the need for businesses to control their AI infrastructure. The article discusses using local models like Gemma and Qwen within an agent harness to perform classification tasks, specifically for triaging issues and pull requests in the OpenClaw repository. This local approach, powered by high-capacity hardware like NVIDIA's GB10, offers near-instantaneous notifications and cost-efficiency, excluding electricity costs. The system employs an agentic classification method, allowing models to search for context before returning structured data. Despite initial challenges with false positives, larger models demonstrated improved precision and recall, offering a viable alternative to cloud-based solutions, particularly for tasks requiring high throughput and quick prototyping. The approach is versatile and applicable to various domains, such as news categorization, social media filtering, and customer support triage, promoting the use of medium-sized local models for efficient and secure classification.
Jun 22, 2026
2,891 words in the original blog post.
V-Zero introduces a novel approach to fine-grained visual reasoning that eliminates the need for annotated answer labels by utilizing a contrastive evidence-gated distillation method. The framework involves a student model that samples reasoning trajectories from the full image, while a teacher model evaluates these trajectories by comparing positive and negative visual evidence views. This contrastive analysis helps determine the degree to which a trajectory is grounded in actual visual evidence, allowing for selective distillation of teacher signals that are more likely to be supported by relevant visual data. By avoiding reliance on language priors and focusing on evidence-based reasoning, V-Zero provides dense token-level supervision during training without altering the standard full-image inference process. The approach leverages on-policy rollouts and evidence-aware distillation to refine model performance in identifying small objects, reading local text, and comparing subtle visual attributes, thus addressing challenges in multimodal large language models without the need for human-written ground-truth answers.
Jun 22, 2026
859 words in the original blog post.
The article explores a novel approach to pretraining medical encoders by leveraging heterogeneous web data instead of traditional hand-curated corpora, which are often limited in scale and diversity, particularly in non-English languages. The proposed methodology utilizes a three-stage data curation pipeline that includes medical-term density filtering, multi-axis annotation, and signal-amplifying rephrasing with large language models (LLMs) to enhance the utility of web-sourced data for medical pretraining. This approach has led to the development of FineMed, a French medical pretraining corpus, and its rephrased subset FineMed-rephrased, alongside the DoctoBERT family of encoders. The evaluation of these encoders demonstrates significant improvements over existing models on a range of French medical natural language processing tasks, highlighting the effectiveness of combining rephrased data with filtered web data to outperform traditional curation methods. The results suggest that this strategy, which taps into the scale and heterogeneity of web data, offers a competitive alternative to narrow, hand-curated corpora, and the study indicates plans to expand the approach to multilingual settings.
Jun 20, 2026
2,019 words in the original blog post.
The multimedia landscape has been transformed by an innovative approach that turns photos into collectible 3D figurines without using traditional creative software like Photoshop or Blender. This process is facilitated by an interactive studio that employs two Hugging Face Spaces, allowing users to generate figurines by simply uploading a photo and using a prompt. Each step in the transformation—stylizing a portrait and reconstructing a 3D mesh—has been simplified into HTTP calls, eliminating the need for complex software integration or expertise. The creative stack has evolved into a system of easily accessible endpoints that agents can chain together rapidly, showcasing a shift from learning and using multiple software applications to leveraging documented API interfaces. This transformation highlights the potential future of multimedia software development, where the focus is on connecting agents to prompts and model endpoints, leading to a more streamlined and accessible creative process.
Jun 19, 2026
1,166 words in the original blog post.
Continuous batching has been introduced as a significant improvement for transformers, specifically in the context of GRPO in TRL, aiming to enhance speed and memory efficiency during training and generation tasks. This advancement provides an in-process solution that fills the gap between the default generate() function and vLLM, eliminating the need for a separate inference engine and weight synchronization between model copies. By using a single flag in the GRPOConfig, users can leverage transformers directly with continuous batching to achieve faster and more resource-efficient rollouts, particularly beneficial for large generation batches with variable completion lengths. Benchmarking on an A100 80GB with Llama-3.2-1B-Instruct demonstrates notable performance improvements, with continuous batching outperforming the default at higher N values. This method is currently text-only and requires transformers version 5.8.0 or later, with ongoing developments promising further enhancements in performance and functionality.
Jun 19, 2026
712 words in the original blog post.
MosaicLeaks highlights the privacy risks associated with deep research agents that combine private local documents with external web tools, posing the threat of sensitive information being inferred from seemingly innocuous web queries. This phenomenon, termed the "mosaic effect," allows adversaries to piece together private information from these queries without direct access to the documents. To address this, MosaicLeaks introduces a deep-research task with multi-hop questions that interleave public and private data. The study reveals that training agents solely for task performance often exacerbates privacy leakage, as more informative queries, while improving task success, lead to higher leakage. To combat this, the Privacy-Aware Deep Research (PA-DR) method is proposed, which combines task performance with privacy considerations, significantly reducing leakage from 34.0% to 9.9% while maintaining high task success rates. The findings emphasize that privacy cannot merely be prompted into agents but must be trained, demonstrating that rewarding careful query construction can drastically cut leakage without compromising effectiveness.
Jun 18, 2026
1,889 words in the original blog post.
The blog post discusses the benchmarking of coding agents working with open models, particularly focusing on transformers, to evaluate not only the correctness of their outputs but also the efficiency of the processes they use to arrive at these results. As coding agents can autonomously select libraries, execute calls, and debug errors, the blog emphasizes the importance of designing software that is not only functional but also agent-friendly, with intuitive APIs and thorough documentation. The study explores how different models and library revisions impact the agent's performance in terms of cost, latency, token usage, and errors, using a tool-specific benchmark for this purpose. It presents the findings that while larger models benefit from a newly introduced CLI and Skill, making task completion faster and more efficient, smaller models struggle with this new interface, leading to increased token consumption and potential decreases in accuracy. The post advocates for testing software specifically for agentic-use to optimize both the tools and the processes for agent interactions, providing insights for library maintainers on improving agentic-optimized tooling.
Jun 18, 2026
3,363 words in the original blog post.
The blog post explores the landscape of parameter-efficient fine-tuning (PEFT) techniques, particularly focusing on the popular Low Rank Adaptation (LoRA) method, which dominates the field due to its early introduction and widespread support. Despite its popularity, the post raises questions about whether it is truly the best choice, as various research papers claim superior performance of alternative PEFT methods. Hugging Face has developed a PEFT library and benchmarking framework to objectively evaluate different techniques under the same conditions, allowing for comparison beyond just test performance to include memory usage and other metrics. The results indicate that while LoRA performs well, other techniques can surpass it in certain aspects, suggesting a need to consider these alternatives based on specific requirements. The post emphasizes that LoRA should not be the default choice and encourages users to explore other PEFT methods supported by the library, taking advantage of the unified API to easily switch between techniques.
Jun 18, 2026
2,754 words in the original blog post.
In a comprehensive benchmarking analysis conducted in June 2026, the enterprise AI tool Falconer consistently outperformed its competitors—Notion, Atlassian Rovo, Claude Code, and Codex—across a variety of retrieval tasks using real-world support and engineering datasets. The evaluation involved 200 questions from two public datasets, including a support corpus and an open-source codebase, with performance judged by advanced models like Claude Opus 4.8 and GPT-5.5. Falconer demonstrated superior capabilities in answering real support and engineering questions, achieving the highest win rates across various head-to-head matchups. The analysis highlighted Falconer's efficient response times and its ability to deliver concise answers, with scoring based on criteria such as faithfulness, helpfulness, completeness, and relevance. The study utilized public and reproducible corpora, ensuring transparency and allowing for re-evaluation by others, while emphasizing that Falconer's advantage was evident even when accounting for different scoring methods and tie rates in the results.
Jun 18, 2026
1,668 words in the original blog post.
GLM-5.2, the latest flagship model from Z.AI, advances long-horizon task capabilities significantly over its predecessor, GLM-5.1, by supporting a stable 1M-token context. The new model introduces several innovations, such as improved architecture through IndexShare to reduce computational costs and an enhanced MTP layer for speculative decoding, thus increasing acceptance length by 20%. It is released under an MIT open-source license, removing regional restrictions and enhancing technical access. GLM-5.2 consistently ranks as the top open-source model across multiple long-horizon coding benchmarks, demonstrating its practical application in sustained engineering work. It also incorporates effort level control, allowing users to balance performance against computational cost, and has been optimized for efficiency in large-scale agentic reinforcement learning tasks, with infrastructure support from the slime framework. An anti-hack module is introduced to mitigate reward hacking in coding scenarios, maintaining training integrity. The model's architecture and enhancements enable it to outperform previous iterations and rival closed-source models in various coding and reasoning benchmarks, while providing users with greater flexibility and scalability in long-context inference scenarios.
Jun 17, 2026
2,853 words in the original blog post.
This article explores the integration of LeRobot in Strands Robots, an open-source SDK from AWS that facilitates the transition from dataset to physical robot across simulation and real-world environments. The integration utilizes AgentTools to create a Strands agent capable of handling robotic tasks with a unified approach, allowing seamless transition between simulation and hardware using the same dataset format. The process involves five steps: building the agent, recording a demonstration in simulation, running a policy, deploying to hardware, and coordinating multiple robots using a peer mesh. The Strands Robots SDK simplifies the robotics workflow by allowing users to compose robotic tasks in natural language, leveraging LeRobot's existing capabilities for hardware interaction and dataset management. The system supports both simulation and real-world environments by maintaining a consistent data format, reducing the complexity of transitioning between these modes. The setup does not require advanced hardware, making it accessible for various skill levels, while emphasizing crucial security considerations for production environments. The article provides a comprehensive guide for building and deploying AI agents for robotics using the SDK, highlighting its potential to democratize access to robotics development.
Jun 17, 2026
3,491 words in the original blog post.
MolmoMotion is an advanced 3D motion forecasting model designed to predict future trajectories of objects in 3D space using video frames, 3D points, and language instructions, outperforming existing methods. By representing motion as object-attached 3D points, the model achieves efficient and accurate predictions across various scenarios, making it useful for applications such as robotics planning and video generation. The model is trained on MolmoMotion-1M, the largest dataset of 3D point trajectories paired with action descriptions, and evaluated using PointMotionBench, a benchmark for measuring 3D motion forecasting accuracy. MolmoMotion employs two variants: autoregressive for smooth and accurate predictions and flow-matching for handling uncertainty, allowing it to adapt to different downstream tasks. Despite its promising capabilities, the model has limitations in handling complex deformable motions due to the sparse query points used during training. Nonetheless, MolmoMotion represents a significant step toward anticipating object movements, offering potential applications beyond perception in fields like robotics and video generation, and encourages further exploration and customization by the community.
Jun 17, 2026
1,901 words in the original blog post.
Xe-Forge is an Intel project designed to optimize Triton kernels for Intel Arc Pro GPUs using a sequence of optimization stages driven by a large language model (LLM). This process, called CoVeR (Chain-of-Verification-and-Refinement), involves a loop that tests and iterates kernel candidates on the GPU to enhance performance. The Xe-Forge framework leverages a knowledge base of Intel XPU-specific patterns to guide optimization, which is often underrepresented in LLM training data. On the Intel Arc Pro B70, Xe-Forge achieves significant speedups over existing PyTorch and Triton kernels, demonstrating its ability to enhance even hand-tuned kernels. The xpu-kernels skill packages this optimization process into an Agent Skill, allowing a coding agent to perform the optimization loop without requiring the entire project. Xe-Forge's effectiveness has been proven across various kernel configurations, particularly in memory and compute-bound scenarios, and it emphasizes the importance of knowledge access in optimizing kernels for less-represented architectures like Intel's XPU.
Jun 17, 2026
2,201 words in the original blog post.
Agentic Resource Discovery (ARD) is an open specification developed collaboratively by industry leaders like Microsoft, Google, and Hugging Face, designed to enhance how AI agents find and connect to tools, skills, and other agents across the web. Unlike the traditional model that requires pre-installation of capabilities, ARD enables dynamic and intent-based search through federated registries, allowing agents to discover and utilize capabilities at runtime. The specification introduces a static manifest format, ai-catalog.json, and a dynamic registry API for live, ranked discovery, facilitating a shift from static catalogs to a more flexible and scalable system. Hugging Face's implementation of ARD demonstrates this approach by allowing its Discover Tool to provide search access to various skills and applications, integrating seamlessly with existing services and supporting different media types. The focus on verification ensures that the capabilities found are trustworthy, emphasizing the importance of verifying publisher identity to prevent unauthorized or tampered interactions.
Jun 17, 2026
1,418 words in the original blog post.
Antoine Chaffin's article explores the challenges and solutions associated with efficient Approximate Nearest Neighbor (ANN) methods for ColBERT models, particularly focusing on issues arising from embedding geometry. While traditional methods like MUVERA and SMVE initially promised to simplify ColBERT infrastructure, their performance faltered with newer models due to embedding anisotropy. Chaffin identifies mean-centering as a partial fix and introduces STE-based regularization, which unexpectedly condenses embeddings into fewer dimensions, enhancing their compatibility with random projections. This regularization technique, which improves performance across various methods without degrading full MaxSim retrieval, suggests that the effective dimensionality of ColBERT spaces is lower than previously assumed. The study not only addresses current inefficiencies but also lays the groundwork for developing more robust and scalable retrieval models and indexing methods.
Jun 16, 2026
5,150 words in the original blog post.
Ishan Awasthi developed an AI Interview Coach for the Gradio × Hugging Face Build Small Hackathon 2026, designed to enhance job interview preparation by transforming unstructured job descriptions into structured role analyses and providing interactive practice sessions with context-aware feedback. The platform utilizes a semantic evaluation methodology, moving away from rigid keyword checking to assess candidates' understanding of technical concepts more naturally. It features a lightweight, decoupled architecture that operates efficiently within strict hackathon constraints, powered by a high-availability cloud serverless API and leveraging tools like Gradio and ReportLab. Awasthi's journey highlighted the importance of stability over refinement under tight deadlines, the flexibility of Gradio, and the benefits of decoupling components to ensure adaptability and ease of troubleshooting. Users can engage with the system by inputting job descriptions, selecting the depth of mock interviews, and receiving detailed feedback, ultimately compiling results into a downloadable PDF report for future reference.
Jun 15, 2026
1,163 words in the original blog post.
Eyas is an AI-powered offline CCTV intelligence agent designed to address the challenges faced by small retail shop owners in monitoring theft in real-time. Developed during the Build Small Hackathon, it leverages a sequence of small models, including YOLO11n for detection and tracking, MiniCPM-V 4.6 for observation, Nemotron 3 Nano 4B for event-log reasoning, TinyAya for Korean translation, and VoxCPM2 for audio synthesis, all running locally without cloud dependencies. The system aims to provide owners with immediate alerts on suspicious activities, allowing them to act promptly rather than relying on footage after incidents occur. The design includes a heuristic structurer that converts video observations into structured event logs, enhancing the efficiency and consistency of the language model's outputs, and it features a custom React frontend for improved user interaction. Despite challenges like the need for GPU for audio synthesis and a limited context window for long recordings, the project demonstrates a significant step towards real-time, autonomous security monitoring for small businesses.
Jun 15, 2026
1,983 words in the original blog post.
Thomas Kim's research explores the impact of QLoRA SFT distillation on the Qwen3.6 27B model, particularly focusing on agentic coding harness fluency via Terminal-Bench 2.0 evaluations. The study investigates how harness-specific fine-tuning can alter model behavior, emphasizing the sensitivity of these changes to training traces, reasoning formats, and harness interfaces. The research tested various harnesses, including Codex CLI, OpenHands, and Pi, finding that the base Qwen3.6 27B model generally performed the best, although the v2 reasoning-distilled model showed improved task decomposition and validation. However, the v2 model also exhibited a tendency to over-explore and time out due to a shorter timeout period compared to the base model. The experiments underscore that while reasoning distillation may enhance certain forms of harness fluency, it can also lead to increased exploratory behavior that may not always be beneficial.
Jun 15, 2026
1,939 words in the original blog post.
"Brad Did Something" is a 2D top-down office comedy game that replaces traditional middle management with a large language model (LLM) to manage chaotic corporate bureaucracy. Developed for the Hugging Face Build Small Hackathon, players assume the role of Head of Sales at Veloura Technologies, navigating unpredictable workplace scenarios to achieve a $1,000,000 revenue goal. The game's innovative mechanics rely on generative AI, with live NPC dialogue and fiscal consequences driven by an LLM, while visual humor is added through AI-generated crisis comics. The game operates on the llama.cpp runtime with Qwen3.5-9B for real-time interaction, ensuring fast, schema-enforced responses that maintain comedic timing. The game tackles challenges like latency and AI behavior, employing server-side validation for consistency and reliability. "Brad Did Something" blends interactive storytelling with AI-driven spontaneity, offering a witty take on corporate culture.
Jun 15, 2026
1,709 words in the original blog post.
OpenMythos, a cybersecurity-focused large language model, was developed in response to the inadequacies of general-purpose LLMs in accurately addressing cybersecurity issues. The model was trained using a meticulously curated dataset that combined formal academic research from the ArXiv cs.CR category with real-world CVE data to provide both theoretical and practical insights into vulnerabilities. The training process involved two stages: Supervised Fine-Tuning (SFT) to establish a foundation for cybersecurity reasoning, followed by Reinforcement Learning with Verifiable Reward (RLVR) to ensure output accuracy by having the model verify its own responses against known vulnerabilities. The training utilized Modal's serverless GPU infrastructure, specifically H100s, to efficiently handle the computational demands without managing long-running instances. The OpenMythos model, along with its datasets and demo, is publicly available on Hugging Face, inviting further evaluation and use in security tooling and vulnerability analysis workflows.
Jun 15, 2026
1,139 words in the original blog post.
In June 2026, Sergio Paniego experimented with a new feature from Google Colab, the Colab CLI, which allows users to fine-tune models without direct interaction with GPUs or writing training scripts. Using the Colab CLI, a coding agent was able to autonomously fine-tune a model by reading examples from the TRL repository, writing a training script, provisioning a GPU, installing dependencies, and managing the training process, including error handling. This process was executed on a free Colab T4, with the training metrics streamed live to a Hugging Face Space, showcasing the model's performance in real-time. The entire operation was cost-free and efficient, highlighting a new era of automation in model training where minimal user input is required, and adjustments can be made by simply altering the initial prompt. The fine-tuned model is stored on the Hugging Face Hub, ready for use, demonstrating the potential for easily adaptable and scalable AI model training using cloud-based tools.
Jun 15, 2026
922 words in the original blog post.
Carbon-VEPor is an innovative system designed to automate Variant Effect Prediction (VEP), which assesses whether genetic mutations are pathogenic or benign by integrating deep biological sequence modeling with rapid deterministic classification. The system employs an autonomous ML-Intern agent utilizing NVIDIA's Nemotron-3-Nano-4B to handle data engineering tasks, from parsing and streaming datasets to generating and modifying scripts for sequence extraction and classification. Key features include computing Log-Likelihood Ratios (LLR) using a Carbon-3B model to measure statistical disruptions caused by mutations and optimizing a neural decision boundary with a 3-layer Multi-Layer Perceptron (MLP) for binary classification. The production pipeline, orchestrated by a central coordinator, executes multi-stage inference to transform clinical PDF reports into structured data, computes LLR scores, and performs classification using recompiled NumPy operations for efficient processing. This end-to-end machine learning approach enhances prediction accuracy and speed, making it a valuable tool for genomic analysis in clinical settings.
Jun 15, 2026
1,743 words in the original blog post.
FINAL-Bench Quantum is a benchmark designed to address the challenges in comparing quantum-computing methods by providing a neutral and standardized platform for evaluation. It features two main tracks: Track A, where methods are assessed on a public test set with verified measurements, and Track B, where results are reported from other sources, acknowledging the differences in codes, noise models, and hardware. The benchmark comprises five events focusing on logical error rates, optimization, molecular energy calculations, quantum-memory query fidelity, and classical simulation capabilities. Each event is categorized into verified measurements, real hardware data, and published references, with detailed charts and medals awarded for participation. FINAL-Bench emphasizes neutrality, ensuring trusted results by faithfully quoting sources and resisting hype, with methods from major companies and institutions evaluated under identical protocols. Submissions are welcome, and a methods paper is in preparation to further advance this initiative.
Jun 14, 2026
679 words in the original blog post.
PitchFight AI is an innovative platform designed to help student founders hone their startup pitches by practicing in a simulated pressure environment before facing real investors or judges. Developed for the Hugging Face Build Small Hackathon, this tool allows users to refine their pitch through structured feedback and a series of challenging questions posed by AI-powered personas, such as skeptical venture capitalists or technical judges. Using NVIDIA Nemotron for its reasoning model, PitchFight AI transforms raw startup ideas into structured presentations and subjects them to rigorous interrogation across multiple rounds, culminating in a scorecard that highlights strengths and areas for improvement. The platform emphasizes the importance of preparing for tough questions, offering a dynamic space where founders can defend their assumptions under pressure, thus enhancing their readiness for real-world pitching scenarios.
Jun 14, 2026
743 words in the original blog post.
Closet Twin is an AI-powered personal stylist developed during the Build Small Hackathon, designed to revolutionize wardrobe management by using computer vision and large language models (LLMs). This intelligent platform addresses the common issue of underutilized wardrobes and style inconsistency by offering a data-driven approach to personal style. With features like digital wardrobe management, AI-powered outfit generation, and personal analytics, users can upload their clothing to receive organized categorization, instant searchability, and personalized outfit recommendations based on various factors such as weather, occasion, and personal style preferences. Additionally, the platform allows users to recreate looks from fashion inspirations and provides insights into their style habits, helping them maximize their clothing investment and build a more cohesive wardrobe. Built with cutting-edge technologies like MiniCPM-V-4.6 for image analysis, Gradio for UI, and FastAPI for backend operations, Closet Twin aims to enhance user experience in fashion tech by demonstrating practical applications of multimodal AI and showing how AI can augment human creativity in styling.
Jun 14, 2026
694 words in the original blog post.
The newly released version of the MTEB leaderboard significantly enhances user experience by addressing previous issues of speed and reliability. Built on a scalable framework using FastAPI and Svelte, the updated leaderboard offers improved filtering, model comparison, and transparency, enabling users to deeply explore and customize benchmarks to suit specific needs. Notably faster than its predecessors, it allows users to filter on domains, language, modality, and individual tasks, and provides transparency by allowing inspection of datasets and task metadata. The update encourages broader improvements across models, not just top performers, by highlighting factors like size, memory usage, and runtime. Users can easily compare models by pinning them for tailored analysis, and can fetch scores locally via CSV or API. The development team encourages user feedback for further enhancements, reflecting a commitment to continuous improvement and community engagement.
Jun 12, 2026
955 words in the original blog post.
Olmo-eval is an advanced evaluation workbench designed to enhance the model development process for large language models (LLMs) by building on the Open Language Model Evaluation Standard (OLMES). Unlike traditional evaluation tools that focus on static benchmarks or sandboxed environments, olmo-eval offers flexibility in defining and implementing new evaluations, allowing researchers to run benchmarks across various model checkpoints and analyze results in detail. It supports agentic and multi-turn evaluation, providing robust analysis tools to discern whether changes in model performance are significant or just noise. While it shares some features with Harbor, another evaluation framework, olmo-eval is specifically tailored for the iterative and dynamic nature of model development, enabling quick adaptation and integration of benchmarks. It allows for reusable components and modular configurations, ensuring that runtime policy and benchmark logic remain distinct. The tool is open for community use, encouraging collaborative improvements and adaptations in the ongoing development of LLMs.
Jun 12, 2026
1,545 words in the original blog post.
Serge is an innovative GitHub-native AI code review tool designed to seamlessly integrate into existing pull request workflows without adding complexity or redundant comment streams. Utilizing an OpenAI-compatible language model, Serge can be invoked with a simple comment to review pull requests and adhere to repository-specific review rules, offering maintainers the flexibility to edit AI-generated reviews before publication. It supports different modes of operation, including a GitHub Action for quick setups, a GitHub App for organization-wide automation, and a staged web app for human-in-the-loop reviews, allowing teams to choose suitable models and configurations. Serge emphasizes security by treating pull request content as untrusted input and ensuring all repository policies are stored on the default branch to prevent unauthorized changes. This tool is open-source, aiming to enhance code review processes while maintaining human oversight and is adaptable to various deployment needs, including containerized setups and scalable infrastructure beyond demo purposes.
Jun 12, 2026
1,443 words in the original blog post.
Lolaby is an AI-powered tool designed to create personalized lullabies for children, aiming to ease the bedtime routine for parents and caregivers. The system allows children to express their interests and fears through drawings or text, which are then transformed into custom lullaby lyrics and music. The AI operates entirely locally, ensuring privacy by not relying on cloud-based services or external data processing. It uses a vision-language model to interpret the drawings, a fine-tuned language model to generate lyrics, and synthesized instruments to compose the music, all designed to provide a calming and comforting experience. The tool emphasizes the importance of specificity and emotional connection, as demonstrated by the positive reactions from a kindergarten teacher who found the personalized songs more engaging than standard lullaby applications. Through careful dataset curation and real-time processing, Lolaby creates a seamless and non-intrusive bedtime interaction that resonates with children by incorporating their personal experiences into the lullabies.
Jun 11, 2026
1,484 words in the original blog post.
In the second part of the "Profiling in PyTorch" series, the authors explore the transition from using basic matrix multiplication and addition to implementing a more sophisticated Multilayer Perceptron (MLP) using the nn.Linear module with the PyTorch profiler. The discussion highlights how nn.Linear wraps matrix multiplication with built-in bias handling, optimizing the process by using epilogues to merge operations and reduce memory traffic. The series progresses by profiling a GeGLU MLP, where the authors observe the benefits of operation fusion using torch.compile, which reduces the number of GPU kernel launches by combining pointwise operations. The exploration also includes a comparison with hand-tuned Triton kernels from the Hugging Face Hub, demonstrating the trade-offs between generic and specialized kernels in terms of performance and flexibility. Through these analyses, the authors emphasize the importance of forming hypotheses before examining profiler traces to effectively diagnose and optimize PyTorch models.
Jun 11, 2026
3,681 words in the original blog post.
PhysicsIntern, initially launched as an autonomous agent to tackle complex physics research problems, has been redesigned to function as a collaborative research assistant rather than a fully independent entity. The original version successfully demonstrated its structured, multi-agent approach by outperforming baseline models on the CritPt benchmark. However, the feedback indicated a preference for a tool that researchers could interact with, guiding and steering the process rather than relying on an autopilot. The revamped PhysicsIntern allows for greater human involvement, acting more as a set of skills integrated with existing coding tools like Claude Code, Codex, or Pi. It utilizes a git-based system to track progress, ensuring transparency and continuity across research sessions. The new version encourages researchers to participate actively, approving plans and engaging with the research process, which allows for dynamic problem-solving and adaptation to complex, open-ended questions. This shift aims to provide researchers with a more flexible, interactive, and efficient research partner that enhances the problem-solving experience.
Jun 11, 2026
2,125 words in the original blog post.
Optimum Intel 2.0 is a major update to the library, now focusing on being an OpenVINO-first toolkit, streamlining the installation process, and providing day-one support for the latest open models. The update simplifies the user experience by removing the Intel Neural Compressor and Intel Extension for PyTorch integrations, previously deprecated, and eliminating the ONNX dependency, while installing OpenVINO and NNCF by default. This version enhances quantization and inference capabilities, allowing efficient deployment of models on Intel hardware such as CPUs, Arc GPUs, and Core Ultra NPUs. Users can benefit from smarter quantization, improved compression, and compatibility with modern architectures, making it suitable for deploying open models on Intel hardware, building on-device or edge AI, and working across various modalities like text, vision-language, speech, and video. The update offers a consistent API for these tasks, making it an appealing option for those looking to efficiently run the newest open models locally on Intel silicon.
Jun 11, 2026
1,038 words in the original blog post.
Reachy Mini, a robot developed by Pollen Robotics, interacts with its surroundings using audio and video processing capabilities facilitated by its Raspberry Pi Camera 3 Wide and custom microphone array. The design allows seamless local and remote usage, with a consistent API for both the Lite and Wireless versions. GStreamer is employed for media handling, enabling direct AI app development using the robot's audio and video streams on various platforms, including laptops and Hugging Face Spaces. WebRTC facilitates low-latency streaming and control, while the SDK simplifies installation and usage. This setup supports advanced applications like speech recognition and object tracking, offering flexibility between local and remote processing. The design ensures minimal latency for real-time interactions, with open-source code available for further development.
Jun 10, 2026
2,694 words in the original blog post.
Wave Function Collapse (WFC) has been harnessed to create "Infinite London," a procedurally generated, endlessly explorable Victorian city built from 36 AI-generated 3D tiles. By leveraging the WFC algorithm, which assembles modular tiles into cohesive arrangements like a sudoku solver, the project transforms text prompts into 3D assets, overcoming the traditionally slow and resource-intensive text-to-3D generation process. The modular tiles, created using ideogram-ai and Microsoft TRELLIS.2, form the building blocks of the city, which can be explored in a browser using WASD controls. The city expands infinitely, with each new world generated from a random seed, offering an ever-changing urban landscape without the need for real-time asset creation. This approach not only simplifies the creation of complex virtual environments but also allows for the sharing and exploration of unique cityscapes through shareable seeds, demonstrating the potential of AI in game development and urban simulation.
Jun 10, 2026
805 words in the original blog post.
In a groundbreaking move, Arcee AI, a prominent American AI lab, has partnered with Hugging Face in a multi-million dollar agreement to utilize Hugging Face Private Storage for all of its models, datasets, and agent traces, both public and private. This collaboration marks a significant shift, as Arcee AI will rely on Hugging Face as the exclusive platform for storing and distributing their AI artifacts, which includes proprietary data and models. The use of Hugging Face's Buckets service allows Arcee AI to maintain compute flexibility without incurring egress fees, supporting the lab's commitment to openness and global accessibility. This partnership highlights Arcee AI's role as a key player in the American open-source AI ecosystem, demonstrating that effective AI solutions can be achieved with specialized models and high-quality data. Hugging Face's infrastructure will facilitate broader distribution and integration of Arcee's work, further bridging the gap between closed and open-source AI development.
Jun 09, 2026
900 words in the original blog post.
In this article, Mishig Davaadorj explores the innovative process of building a 3D gallery of Parisian monuments by chaining two Hugging Face Spaces, highlighting a shift towards a "building block economy" in software development. By using a coding agent, Davaadorj demonstrates how multimedia assets can be efficiently created and integrated by leveraging small, well-documented components from the Hugging Face Hub. The process involved using the ideogram-ai/ideogram4 Space to generate images of monuments and the VAST-AI/TripoSplat Space to convert those images into 3D Gaussian splats, without the need for traditional image generation or 3D reconstruction tools. The key to this seamless integration is the use of agents.md documentation, which simplifies the process of chaining different models and eliminates the complexities of integration. This method exemplifies how agents can now easily compose state-of-the-art models into functional applications, suggesting a future where software development is more about assembling existing components than building from scratch.
Jun 09, 2026
907 words in the original blog post.
The exploration of how voice agents handle code-switched speech reveals critical insights into the performance of automatic speech recognition (ASR) systems when dealing with bilingual customers who naturally switch languages. The benchmark study focuses on four language pairs—Spanish-English, French-English, Canadian French-English, and German-English—in enterprise settings like HR and IT scenarios, assessing models through metrics such as Word Error Rate (WER), Semantic Word Error Rate (SWER), and Answer Error Rate (AER). The study finds that transcription accuracy and semantic understanding vary significantly across models, with ElevenLabs Scribe V2, Gemini 3 Flash, and AssemblyAI Universal 3-Pro leading in performance. It highlights that code-switching introduces varied challenges depending on the language pair and context, exposing differences in model robustness rather than uniformly increasing difficulty. The number of language switches within an utterance is a key factor in transcription errors, while the Code-Mixing Index (CMI) influences error magnitude. Interestingly, errors predominantly occur in the English segments of code-switched utterances despite English being well-handled in monolingual contexts, suggesting that embedded language segments pose additional transcription challenges. This study underscores the importance of benchmarking ASR systems against the specific language pairs relevant to an enterprise's customer base to ensure effective handling of code-switched speech.
Jun 09, 2026
2,621 words in the original blog post.
Abubakar Abid outlines a method for migrating GitHub CI workflows to Hugging Face Jobs, highlighting the benefits of using Hugging Face's serverless infrastructure for Continuous Integration (CI) tasks. This approach addresses the limitations of GitHub Actions, such as slow processing and lack of GPU access, by utilizing Hugging Face Jobs to run commands or scripts on various hardware configurations, including GPUs. The integration involves creating a GitHub App and a dispatcher Space to connect GitHub Actions with Hugging Face Jobs, allowing CI jobs to run on optimized hardware and streamlining the process of executing both CPU and GPU tests. The migration significantly reduces CI time and costs for projects like Trackio, and provides flexibility in choosing Docker images and utilizing advanced features like mounting volumes, thereby enhancing the efficiency and scope of CI workflows.
Jun 09, 2026
1,753 words in the original blog post.
Cohere has introduced North Mini Code, a 30B-parameter Mixture-of-Experts model optimized for software engineering tasks, available on Hugging Face under the Apache 2.0 license. This model, the first in a new family, is designed for complex coding workflows and outperforms larger models in agentic coding benchmarks. North Mini Code employs a Mixture-of-Experts Transformer architecture and undergoes a rigorous training process involving supervised fine-tuning and reinforcement learning with verifiable rewards, focusing on agentic coding tasks. The model's training includes a diverse data mixture to enhance robustness across various coding harnesses and environments, improving performance on benchmarks like SWE-Bench and Terminal-Bench. North Mini Code also benefits from asynchronous reinforcement learning to optimize agentic coding rollouts, and it shows notable improvements in robustness and efficiency over its SFT-only counterpart, particularly in code editing tasks. The model is accessible in OpenCode, Cohere API, and on Hugging Face with both BF16 and FP8 weights.
Jun 09, 2026
2,737 words in the original blog post.
NeuroBait is an AI project developed to assist individuals, particularly those with ADHD, by sparking dopamine to help initiate tasks rather than relying on traditional to-do lists that often exacerbate feelings of being overwhelmed. Created by Harisabekti Dicky Subrata as a personal initiative inspired by his wife's struggles with ADHD, NeuroBait is designed to address executive dysfunction and task-initiation paralysis by focusing on real-life observations and practices. It identifies what truly matters to the user, such as a deadline or a personal interest, and provides a gentle nudge in the form of a small, actionable step without guilt or lectures. The model, fine-tuned from Google’s Gemma-3-12b, is implemented via Hugging Face Space and offers responses in warm, flowing language tailored to the individual, making it applicable not only to ADHD sufferers but to anyone feeling overwhelmed by modern life’s demands. The project's future aims include expanding its accessibility, offering bilingual support, and incorporating community feedback to refine its utility.
Jun 09, 2026
794 words in the original blog post.
Pakistan Notice Helper is an AI tool developed for the Hugging Face Build Small Hackathon to address a local safety issue in Pakistan by assisting users in identifying suspicious messages that appear to come from legitimate sources like banks or government departments. Rather than determining the authenticity of messages, it functions as a triage tool, providing users with risk assessments, explanations, visible warning signs, and recommended next steps in both English and Urdu. The project emphasizes the effectiveness of small, focused AI models, with Qwen3.5 4B chosen for its balance of quality, speed, cost, and feasibility over larger models. The tool is designed to be a practical, local solution, helping users recognize red flags and avoid potentially fraudulent actions. Future improvements aim to introduce a verification workflow to further assist users in distinguishing between genuine and fraudulent messages. The project highlights the power of small AI models when applied to narrowly defined, specific problems, providing a useful service to its target audience while acknowledging its limitations.
Jun 08, 2026
2,724 words in the original blog post.
OpenEnv is an open-source tool designed to create agentic execution environments, allowing agents to interact with various interfaces such as terminals and browsers. The initiative, now hosted on Hugging Face and backed by top organizations like Meta-PyTorch, Nvidia, and Hugging Face, aims to standardize the interface between harnesses, environments, and trainers to train more effective open-source models. It acts as an interoperability layer for reinforcement learning environments, promoting efficiency by enabling a one-interface-fits-all approach across different ecosystems without imposing on reward definitions or training specifics. The project's future developments include integrating task sets with Hugging Face datasets, supporting external rewards, enhancing harness integration, providing end-to-end training examples, and establishing auto-validation for environment quality. OpenEnv invites community involvement to refine and expand its capabilities, emphasizing its role as a foundational component for open-source agentic reinforcement learning.
Jun 08, 2026
850 words in the original blog post.
Her, a tool named after the Marathi word for "detective," is designed to analyze and decipher the intricate details of Claude Code sessions by reading the .jsonl files generated during these sessions. The tool processes these files to provide a clear, English-language summary of session activities, identifying potential issues and suggesting improvements based on best practices from Anthropic and the community. Her operates without relying on third-party AI APIs, ensuring privacy by using a deterministic evaluation engine and the Nemotron-Mini-4B-Instruct model for generating suggestions. It highlights key activities like tool deployments and production changes, offering users insight into the usage of various tools and subagents. The tool was initially developed over a weekend to visualize query costs and has since evolved to include an executive report and a comprehensive database for identifying command-line tools used in sessions. Her is accessible on Hugging Face, running on its own GPU to maintain user privacy.
Jun 07, 2026
622 words in the original blog post.
The article presents a detailed account of assembling and upgrading a LeKiwi mobile manipulator integrated with a PincOpen parallel gripper, focusing on resolving technical challenges and optimizing performance for real-world tasks. The author shares insights on overcoming issues such as servo failures during forward stretches by implementing a 3D-printed support block, adjusting P-gain settings, and upgrading to stronger STS3250 servos. The integration with LeRobot 0.5.2 is highlighted, emphasizing modifications to support the LeKiwi system, including a dual-teleoperation setup and customized camera configurations. The article also explores the impact of data scaling, inference modes, and fine-tuning on task performance, revealing that unfreezing the vision language model (VLM) improved navigation capabilities. The author concludes by suggesting the use of torchcodec for video processing and hints at future explorations in multi-task learning and reinforcement learning to enhance robotic capabilities.
Jun 07, 2026
3,431 words in the original blog post.
Job hunting for new graduates can be an overwhelming task, but a newly developed system aims to streamline the process by using AI to generate a shortlist of job opportunities with detailed reasoning for each match. This framework involves a three-step process where a model reads a candidate's resume and preferences to create LinkedIn-shaped search queries, which are then executed to return job postings. These postings are scored on five dimensions: skills match, experience relevance, education and certifications, industry/domain fit, and seniority alignment. The system employs DeepSeek V4 Pro as a "teacher" to generate labels and Qwen3-8B as a "student" to absorb these judgments, resulting in a curated dataset of resumes and job postings. The training process uses two separate LoRA runs to improve performance, and the model operates on a HuggingFace ZeroGPU Space, allowing for efficient job fit evaluations. This innovative approach not only reduces the time and effort involved in job searching but also provides transparency and defensible reasoning behind each job match.
Jun 06, 2026
872 words in the original blog post.
The article provides an in-depth look at the development of "Thousand Token Wood v2," a finance-focused game emerging from the Build Small Hackathon, where each character operates on a different lab's small AI model. Unlike its predecessor, which was more of a passive simulation, the updated version allows players to actively engage as a financier, manipulating the economy through lending, market speculation, and alliance formation while evading a magistrate's scrutiny. The game utilizes four distinct models from different labs to create genuine diversity in character behavior, emphasizing the importance of heterogeneity for an engaging experience. The core challenges in development centered around serving layer friction and ensuring the security of insider information, which was addressed by separating sensitive data from model prompts. The game's persistent memory feature allows characters to develop relationships based on past interactions, enhancing realism without overwhelming the models. The article concludes by highlighting the potential of small models for generating complex, interactive environments when combined with robust configuration and testing strategies.
Jun 06, 2026
1,141 words in the original blog post.
The guide provides a comprehensive walkthrough for deploying and operating Anthropic's Claude Code and the open-source OpenCode on Dell Enterprise Hub (DEH), allowing users to run advanced AI models on-premises with full data sovereignty. It emphasizes the benefits of using DEH, such as predictable latency, cost-efficiency by paying for GPU usage instead of per token, and enhanced model control with the ability to fine-tune and audit versions. The guide outlines a step-by-step process to set up these models, detailing necessary configurations and commands for both Claude Code and OpenCode, with each having distinct configuration files and setup instructions. It also explores various frontier open-weight models suitable for coding tasks, available through DEH, and addresses common troubleshooting issues related to model deployment. Ultimately, the guide highlights how running these models locally on Dell PowerEdge platforms ensures that data remains within the data center, offering a robust, customizable, and secure environment for AI-driven software engineering tasks.
Jun 06, 2026
1,723 words in the original blog post.
Thousand Token Wood is a project developed for the Build Small Hackathon, featuring a simulated economy where five woodland creatures act as agents within a 3-billion-parameter model, Qwen2.5-3B, to trade goods, manage resources, and respond to market dynamics. The simulation, accessible via a Gradio app and utilizing vLLM on Modal, demonstrates the challenges and successes of using small models for real-time multi-agent systems. Initially plagued by issues such as self-sufficiency and a lack of trading incentives, the simulation was improved by introducing engineered scarcity through mechanisms like diet variety, spoilage, and a winter fuel crisis, which created market pressures and wealth disparities among the agents. Despite the model's limitations in economic reasoning, enhancements through structured prompts enabled more realistic trading behaviors, while the incorporation of historical market events as "Wood Legends" added dynamic storytelling elements. These adjustments showcase the potential of small models in simulating complex systems, emphasizing the importance of thoughtful design and scarcity to generate engaging and emergent economic interactions.
Jun 05, 2026
1,029 words in the original blog post.
NVIDIA's Nemotron 3.5 ASR is a cutting-edge, multilingual speech-to-text model that transcribes audio in real-time across 40 language-locales using a single 600 million-parameter checkpoint, with built-in punctuation and capitalization. It overcomes traditional challenges in multilingual speech recognition, such as the need for multiple models or APIs, high latency, and lack of language flexibility, by employing a Cache-Aware FastConformer-RNNT architecture that reduces redundant computations. This model is available as open weights on Hugging Face, allowing users to inspect, fine-tune, and deploy it without additional API dependencies. Fine-tuning the model can significantly improve its accuracy for specific languages, domains, or accents, particularly benefiting languages with less pretraining data. The model's flexibility extends to various applications, including sub-second voice agents, live multilingual meeting captions, and on-device transcription, making it a versatile tool for building multilingual speech applications.
Jun 04, 2026
2,254 words in the original blog post.
In this article, Stephen Batifol outlines a step-by-step guide for fine-tuning the FLUX.2 [klein] model using a LoRA in under an hour on a consumer GPU, specifically for the Build Small Hackathon hosted by Gradio and Hugging Face. The guide emphasizes the use of the 4B model, which fits within 24 GB of VRAM and costs around $0.50 if rented, and details the process of building a dataset, configuring a trainer, and running the fine-tuning loop to create a .safetensors LoRA that imparts a specific style or behavior to the model. It also explains the benefits of using the base model for training and the distilled model for inference due to its faster performance and better results. Additionally, the article covers creating both style and edit LoRAs, where the former involves content-only captions and the latter requires paired datasets, to allow for creative flexibility and tailored outcomes. Finally, it guides users on wrapping the trained model in a Gradio app for deployment as a Hugging Face Space, offering a practical and accessible approach to participating in the hackathon and beyond.
Jun 04, 2026
2,617 words in the original blog post.
EVA-Bench Data 2.0 is an expansive framework designed to evaluate voice agents across three enterprise domains: Airline Customer Service Management, Enterprise IT Service Management, and Healthcare HR Service Delivery, covering 213 scenarios with 121 tools. This release, which represents a fourfold increase in scenario coverage from its original version, aims to test the adaptability of voice agents to varying domain-specific challenges such as vocabulary, workflow complexity, and user expectations. EVA-Bench is validated for solvability against advanced language models like OpenAI GPT-5.4, Google Gemini 3.1 Pro, and Anthropic Claude Opus 4.6, ensuring a robust and fair evaluation process. The dataset is structured to reflect realistic call patterns, including single and multi-intent scenarios, and adversarial interactions, with an emphasis on authentication and reproducibility. Scenarios are generated and validated using a graph-based pipeline to ensure consistency, and the datasets are open-source with multilingual support to address the challenges of deploying voice agents in diverse linguistic environments.
Jun 04, 2026
1,990 words in the original blog post.
In the development of large-scale language models, the focus has shifted from merely the volume of data to the quality and structure of learning signals within the data. Task-seeded synthetic Q&A generation is highlighted as a method that enhances model training by providing structured, task-aligned examples that include clear information needs and enriched explanations. A 100B-token continuation experiment on the Nemotron-3 Nano model demonstrated that this approach improved performance metrics such as MMLU-Pro, average code, commonsense understanding, and GPQA while maintaining stability in average math performance. The process involves using public task training splits as seeds, generating new task-aligned examples, and enriching them with reasoning and knowledge, which are then filtered into curated datasets. These datasets are used in downstream training to enhance model capabilities without overfitting to specific data sources, facilitating positive transfer learning across task families. This methodology not only targets skills crucial for late-stage model training but also ensures that improvements in specific evaluations do not compromise overall knowledge retention.
Jun 04, 2026
1,811 words in the original blog post.
The Hugging Face CLI (hf CLI) has been redesigned to cater to both human users and coding agents, facilitating interaction with the Hugging Face Hub through efficient command-line operations. The CLI enables a range of tasks such as downloading and uploading models, managing repositories, and running jobs, and its recent updates optimize the output for coding agents like Claude Code and Codex by using environment variables to tailor responses. The CLI's efficiency is demonstrated through benchmarks showing it uses significantly fewer tokens than alternatives like curl or the Python SDK on complex tasks, thus improving performance for tasks involving multiple steps. It also introduces agent-mode outputs and hints to direct subsequent actions, making it suitable for both human and agent interactions. Additionally, the CLI includes features like non-blocking commands, predictable command structures, and a "skill" option that provides a compact reference for command usage, which agents can utilize to improve efficiency. As agents become more prevalent in using the Hub, the hf CLI represents a robust tool that streamlines processes for both agents and their human counterparts.
Jun 04, 2026
2,856 words in the original blog post.
NVIDIA's Nemotron 3.5 Content Safety represents an advanced evolution in multimodal and multilingual AI safety models, integrating customizable enterprise policy enforcement and auditable reasoning into a single inference model. Released in June 2026, the model builds on its predecessor by enhancing multimodal integration, allowing simultaneous evaluation of user prompts, images, and responses to identify policy violations that arise from the interplay of different content types. It maintains explicit training coverage across 12 languages while supporting zero-shot generalization to approximately 140 languages, facilitating deployment in regions with limited training data. A significant innovation of Nemotron 3.5 is its custom policy enforcement, enabling domain-specific safety policies across varied applications, from healthcare to education, with a reasoning trace function that provides step-by-step logic for audit and compliance purposes. Despite the potential latency introduced by detailed reasoning processes, the model offers a low-latency binary verdict option and remains efficient, reducing costs and maintaining high accuracy across multilingual and multimodal benchmarks. The release includes a safety dataset, addressing gaps in existing evaluation infrastructure by incorporating real images and culturally nuanced multilingual prompts, and is available under the NVIDIA Open Model License, with support for various inference platforms and tools for custom policy generation.
Jun 04, 2026
2,226 words in the original blog post.
The guide provides detailed instructions for participants in the Build Small Hackathon, emphasizing the use of Cohere's small open models, such as Tiny Aya and Cohere Transcribe, which are designed to operate within a 32 billion parameter limit and are suitable for multilingual text generation and automatic speech recognition, respectively. Tiny Aya supports over 70 languages and is ideal for creating local multilingual assistants and small applications, while Cohere Transcribe covers 14 languages, making it apt for voice interfaces and accessibility tools. The guide offers step-by-step instructions for setting up a Gradio app on Hugging Face Spaces, utilizing tools like llama.cpp for local server setup, and deploying models for various regional language needs. It also discusses the integration of Gradio for building custom user interfaces and demonstrates how to run these models locally or via cloud services, providing insights into model deployment, quantization options, and potential use cases for both text and audio processing applications.
Jun 04, 2026
2,018 words in the original blog post.
An experiment conducted with the support of JarvisLabs explored the nuances of image generation using Diffusers, focusing on the iterative process beyond initial prompt generation. The study utilized Codex as a research agent to manage the selection and testing of variables such as aspect ratio and floor separation in a text-to-image task, which aimed to create a complex architectural watercolor. The experiment highlighted the importance of understanding failures and making adjustments, rather than relying solely on prompts. Successful outcomes resulted from identifying and modifying specific elements like semantic anchors and floor structure, while less effective changes, such as increasing image density or relying on negative constraints, proved informative in understanding the limitations of prompt-based control. The experiment emphasized a modular approach to image generation, leveraging JarvisLabs' GPU capabilities for rapid iteration and emphasizing the value of a structured, agent-driven research methodology over simple prompt manipulation.
Jun 03, 2026
2,294 words in the original blog post.
Reachy Mini, a conversational robot application, has integrated a new feature allowing it to utilize tools hosted on public Hugging Face Spaces via the MCP protocol, expanding its capabilities without requiring direct code modifications. This enhancement enables the robot to perform tasks such as checking the weather or conducting web searches by simply adding a tool from the Hub with a single command, which runs remotely without downloading on the local machine. Built-in tools remain local and focus on the robot’s physical interactions, whereas remote tools, such as the newly introduced weather and search tools, can be added and enabled in specific profiles for enhanced functionality. These profiles control tool usage, with instructions (prompts) guiding the model on how to employ them efficiently during interactions. While local tools remain crucial for operations tied to the robot's hardware, remote tools offer a flexible and shareable solution for non-physical tasks, encouraging innovation and collaboration within the community.
Jun 03, 2026
2,068 words in the original blog post.
DharmaOCR, a specialized OCR model for Brazilian Portuguese text, has demonstrated significant improvements in reducing text degeneration rates through a method called Direct Preference Optimization (DPO). The approach leverages the model's own degenerate outputs as negative training signals, using them to create preference pairs that help the model learn to avoid failure modes. Unlike traditional supervised fine-tuning (SFT), which optimizes for correct predictions without explicitly penalizing degeneration, DPO targets this specific failure mode by training on the full output, rather than token-level predictions. The study showed that DPO reduced degeneration rates across various model families by an average of 59.4%, with some models experiencing reductions as high as 87.6%. This methodology suggests that structured generation tasks like OCR can benefit from using model failures as training signals, provided the failures are distinct, scoreable, and numerous, thereby enhancing model reliability without compromising extraction quality.
Jun 03, 2026
2,953 words in the original blog post.
Holo3.1, an advancement of the Holo3 computer-use model, is designed to enhance robustness across various environments, agent frameworks, and deployment targets, offering seamless integration for both desktop and mobile applications. The release includes quantized checkpoints optimized for local inference, such as FP8, Q4 GGUF, and NVFP4, which support fast and efficient local execution without significant performance loss. With improvements in cross-harness performance and on mobile platforms, Holo3.1 models, ranging from ultra-lightweight to state-of-the-art, cater to diverse deployment needs by balancing cost, performance, and privacy. The new release aims to support flexible deployment options, enabling users to run agents locally on consumer hardware while ensuring data remains private. These developments mark a significant step toward realizing universal computer-use agents that can operate across different settings and devices, and Holo3.1 is now available for developers and enterprises to integrate into their workflows.
Jun 02, 2026
867 words in the original blog post.
ClawHub Security Signals is a dataset comprising 67,453 public agent skills from the ClawHub registry, designed to aid research on agent supply-chain security and multi-signal triage. It integrates data from three scanner families—VirusTotal, static heuristic analysis, and NVIDIA SkillSpector—to produce registry verdicts without human annotations. The dataset reveals significant disagreement among scanners, highlighting the need for ensemble approaches to assess malware reputation, static patterns, and semantic risks associated with agent skills. SkillSpector, with a broader scope, often identifies advisory signals regarding authority and data flow, while VirusTotal excels in detecting malicious content. The dataset is structured into four splits for training, validation, testing, and evaluation, providing sanitized data and redacted sensitive information. It aims to facilitate the development of safe agentic systems by examining scanner disagreements and advancing research areas like multi-signal triage, prompt-injection detection, and least-privilege policy learning.
Jun 01, 2026
1,400 words in the original blog post.
NVIDIA Cosmos 3, now available on Hugging Face, represents a significant advancement in world foundation models (WFMs) for physical AI by offering a unified omni-model that integrates world generation, physical reasoning, and action generation. Built on a Mixture-of-Transformers architecture, Cosmos 3 consolidates capabilities previously handled by separate models, enabling the generation of realistic video worlds, reasoning about physical properties, and predicting future sequences within a single model. It is designed for applications in robotics, autonomous vehicles, and smart spaces, leveraging its ability to understand and simulate complex physical environments. The model comes in two versions—Cosmos 3 Nano, optimized for efficient inference, and Cosmos 3 Super, intended for large-scale synthetic data generation and research. With integration into the Hugging Face Diffusers library, Cosmos 3 facilitates seamless adoption within existing pipelines and supports various input and output modalities. Accompanying the launch are Synthetic Data Generation datasets and resources for post-training, further enhancing its utility for training and evaluating physical AI systems.
Jun 01, 2026
1,960 words in the original blog post.
Scalable enterprise AI adoption hinges on the integration of agent logic, which acts as an intelligent guide to improve agent quality and cost-effectiveness while fostering user trust. This approach addresses the limitations of traditional AI pilots and emphasizes the importance of AI operating at the core of enterprise workflows. By incorporating agent logic—such as knowledge graphs, program analysis libraries, and algorithms—AI systems can achieve more efficient and accurate outcomes, particularly in complex domains like legacy code analysis, test generation, incident response, and compliance modernization. Case studies demonstrate how agent logic can significantly reduce token consumption and improve performance across various enterprise tasks, including healthcare and maintenance of physical assets. This paradigm shift is crucial for realizing the full potential of AI in transforming industries and achieving scalable adoption at optimal costs.
Jun 01, 2026
2,177 words in the original blog post.
Abhishek Sensharma introduces a deep neural network model called Supercut, developed at lucidml, which can transform any image into a playable game in real-time on consumer GPUs, specifically using an RTX 5090 machine. The model, based on the 420M image DiT model from lucidml, incorporates temporal mixing modules and is trained with video and gameplay data to simulate interactive worlds without relying on images from its training dataset. The project operates under a modest compute budget compared to frontier labs, and the current version represents only a fraction of its potential, with an upcoming 800M model expected to enhance motion quality and diversity. The video accompanying the project showcases real play sessions and highlights the innovative use of consumer-level hardware for complex neural network tasks, marking a significant step in generative gaming and world modeling.
Jun 01, 2026
365 words in the original blog post.
JetBrains has introduced Mellum2, a 12-billion parameter Mixture-of-Experts (MoE) model, designed to efficiently handle natural language and code tasks by activating only 2.5 billion parameters per token for high-throughput, low-latency inference. Released under the Apache 2.0 license, Mellum2 is optimized for various applications, including routing, retrieval-augmented generation (RAG), summarization, and private deployments, particularly in software engineering contexts. The model stands out for its competitive benchmark performance and more than double the inference speed compared to similar-sized models, making it suitable for high-frequency tasks within larger AI systems. Mellum2's architecture, which focuses on text and code rather than multimodal tasks, aims to enhance efficiency and reduce costs for real-time workloads, positioning it as a key component in modern AI systems that require specialized yet integrated model components. The model is available for download on Hugging Face, and detailed information on its architecture, training, and evaluation is provided in a technical report.
Jun 01, 2026
600 words in the original blog post.
NVIDIA has announced a major update to its Alpamayo open platform, which is designed to aid in the development of reasoning-based autonomous vehicles (AVs). The platform offers a flexible suite of models, datasets, simulations, and training tools, which have been rapidly adopted across industry and academia, with over 400,000 downloads. The latest iteration, Alpamayo 2 Super, significantly enhances the capabilities of its reasoning models by scaling to 32 billion parameters and introducing advanced features like surround-view camera inputs and meta-action outputs. Additionally, NVIDIA introduced AlpaGym, a closed-loop reinforcement learning framework, which allows for continuous decision-making and learning from experiential feedback. This update also includes new benchmarks to evaluate AV models in realistic scenarios, fostering progress in the field. Alpamayo Recipes serves as a centralized hub for developers to build and customize AV applications using Alpamayo models, with a focus on bridging the gap between training and real-world deployment.
Jun 01, 2026
1,386 words in the original blog post.
The integration of Hugging Face Datasets with Dagster, through the newly introduced dagster-hf-datasets library, bridges the gap between dataset storage and production-scale data pipelines by treating datasets as first-class assets in modern machine learning systems. This integration allows for the orchestration of datasets as evolving operational assets, enhancing their usability beyond static repositories. By modeling datasets as Dagster assets, the library facilitates incremental transformations, metadata tracking, and lineage visualization, enabling datasets to be managed as observable and reproducible entities throughout their lifecycle. This approach not only supports scheduled refreshes and feature extraction but also automates the generation of dataset documentation, making curated datasets more transparent and accessible. The dagster-hf-datasets library underscores the shift towards operational complexity in ML systems, where datasets require orchestration and observability, akin to traditional software infrastructure, to support scalable and trustworthy data workflows.
Jun 01, 2026
1,357 words in the original blog post.