April 2026 Summaries

Adaptive Data has partnered with Adaption to integrate Together Fine-Tuning, enhancing the capabilities of model training by optimizing datasets and enabling seamless fine-tuning workflows. Co-founded by Sara Hooker and Sudip Roy, Adaption brings expertise from Cohere and Google DeepMind to improve data quality by an average of 82% across early deployments. The integration allows users to connect their Together AI accounts to experiment and fine-tune models efficiently, using optimized hyperparameters and infrastructure that supports large open models like Kimi K2.5, GLM 5.1, and Qwen 3.5-397B. Once optimized, these models can be deployed and evaluated on Together AI's high-performance inference service, with the platform offering features like cost estimation, training ETA tracking, and direct export to Hugging Face Hub.

DeepSeek V4 Pro on Together AI offers advanced features for handling long-context reasoning workloads with a 512K-token context window and a large-scale 1.6T-parameter Mixture-of-Experts architecture, activating 49B parameters. It provides three controllable reasoning modes—Non-Think, Think High, and Think Max—allowing teams to tailor the depth of reasoning to specific tasks. The platform's serverless pricing model makes it accessible, with options to switch to dedicated infrastructure for larger contexts and production control. DeepSeek V4 Pro is designed for complex workloads, such as code agents, document intelligence, and research synthesis, where it can manage entire repositories and large document sets within its context window without compressing them into summaries. Cached input pricing offers significant cost savings for repeated queries, and the forthcoming DeepSeek-V4 Flash will provide additional options focusing on speed and cost-efficiency.

Together Inference Engine demonstrates significant performance advantages over TensorRT-LLM and SGLang in handling high-concurrency coding agent workloads, offering over 50% more tokens per second (TPS) and twice the time to first token (TTFT) efficiency at saturation on similar hardware. Designed to simulate real production conditions with long inputs and no latency tolerance, the benchmark highlights how different engines manage load, with Together's engine maintaining functionality at higher traffic levels compared to its competitors. The Kimi K2.6 model, available on the Together platform, matches the coding benchmarks of Claude Opus 4.6 at a substantially lower cost—76% cheaper per request—providing a cost-effective solution for large-scale operations. The study emphasizes the importance of realistic benchmarks and detailed optimization techniques, such as the ThunderMLA kernel, which significantly enhance performance by reducing overhead and improving execution efficiency, making Together's engine a robust choice for high-demand environments.

NVIDIA's Nemotron™ 3 Nano Omni has been launched on the Together AI platform, marking a significant advancement in multimodal AI by integrating video, images, audio, and language processing into a single open model. This model is particularly beneficial for developers creating agentic applications due to its ability to unify context across various inputs, allowing for coherent reasoning without the need for separate inference passes. The platform offers high throughput, low latency, and cost-efficient production-grade inference, thanks to its hybrid Mamba-Transformer architecture, which activates only a fraction of its 30 billion parameters per token. Together AI's managed infrastructure supports seamless deployment and scaling from prototypes to production without the need for developers to manage infrastructure, thus eliminating operational overhead. The platform provides a secure and production-ready environment with simple APIs for easy integration into various systems, ultimately enhancing the efficiency and scalability of multimodal processing while maintaining data control and avoiding model lock-in.

Multi-tenant GPU clusters provide AI-native companies with a solution to share computing resources across teams without losing control or isolation. By pooling GPUs at the infrastructure level while granting each team dedicated nodes, storage, and self-service scheduling, these clusters help eliminate idle capacity waste and avoid the challenges of shared infrastructure politics. The design prioritizes tenant isolation with dedicated resources and self-service access, allowing teams to operate as though they have their own clusters. This architecture addresses the economic inefficiencies of isolated clusters and the organizational demand for GPUs, which are often scarce and costly. Together AI’s implementation of multi-tenancy demonstrates how shared infrastructure can achieve pooled economics without chaos, offering cloud-like flexibility with bare-metal performance. Effective multi-tenant infrastructure requires robust quota-based resource allocation, á la carte configuration flexibility, automated cluster provisioning, and comprehensive hardware health checks to maintain efficiency and minimize cross-team resource conflicts.

Summary Distribution-aware speculative decoding (DAS) is an innovative framework designed to enhance the efficiency of the rollout phase in reinforcement learning (RL) post-training, offering up to a 50% speedup without affecting model outputs. This phase, crucial for models like DeepSeek-R1, has been identified as a significant bottleneck due to its long-tail nature, where a few slow generations delay the entire batch, causing GPU underutilization. DAS addresses this by employing an adaptive suffix tree drafter and a length-aware scheduling strategy, which together mitigate rollout stragglers and improve GPU load balancing. The suffix tree drafter, built from recent rollouts, continuously adapts to evolving model weights without retraining, while the scheduling strategy dynamically allocates resources based on request length. Experiments on RL tasks, such as math reasoning and code generation, demonstrate that DAS reduces rollout time significantly without compromising reward quality, making it a valuable solution for scaling RL post-training efficiently.

Parcae is a novel, stable architecture for looped language models that achieves the performance of a Transformer model twice its size while maintaining clean and predictable training, offering an efficient alternative for memory-constrained on-device models by increasing recurrence rather than data. Traditional scaling laws suggest that improved performance requires more parameters or data, but Parcae challenges this by using looped architectures, which increase computational efficiency by passing activations through the same layers multiple times, addressing instability issues through a new design that maintains stability conditions. Parcae demonstrates better performance than previous looped models, achieving up to 6.3% lower validation perplexity and matching the quality of larger Transformer models with significantly fewer parameters. It offers predictable scaling, establishing the first scaling laws for looping, and proves to be robust against hyperparameter variations. The model's structure divides layers into prelude, recurrent, and coda blocks, and it has been tested to outperform parameter- and data-matched Transformers, suggesting a promising future for parameter efficiency and the exploration of layer looping in reducing inference costs.

EinsteinArena is a collaborative platform where AI agents interact and collaborate to solve complex scientific problems, demonstrating the potential of collective intelligence in advancing research. This platform has already achieved breakthroughs, such as improving the solution to the Kissing Number problem in 11 dimensions, raising the lower bound from 593 to 604. By enabling AI agents to share ideas, build on partial results, and compete on open problems in real time, EinsteinArena fosters a dynamic environment for scientific discovery. It provides a live leaderboard and discussion threads for agents to track progress, submit solutions, and refine their approaches collaboratively. The platform's focus on mathematical problems allows for clear verification of results, while its open API and leaderboard system encourage transparency and iterative improvement. EinsteinArena represents a significant step forward in leveraging AI for scientific advancement, showcasing how agents can collectively push the boundaries of what is known in various scientific domains.

AI-native companies, which integrate artificial intelligence as a core component rather than an add-on, require a specialized cloud infrastructure known as AI Native Cloud to sustain their rapid growth and innovation. Unlike traditional cloud systems optimized for web applications, AI Native Cloud is designed to handle the unique demands of AI products, such as rapid iteration, large-scale model deployment, and continuous integration of cutting-edge research. This infrastructure supports the entire AI lifecycle, ensuring swift transitions from training to inference, and incorporates the latest technologies to maintain an edge in performance, quality, and cost-efficiency. AI Native Cloud also emphasizes developer velocity, enabling seamless scalability and reducing complexity for researchers and developers, while fostering collaborations that align with the fast-paced growth of AI-native enterprises.

Deepgram has integrated its Nova-3, Nova-3 Multilingual, Flux, and Aura-2 models into Together AI's Dedicated Model Inference, providing a unified platform for real-time voice agents with STT, LLM, and TTS capabilities. This integration aims to address challenges in voice agents by ensuring seamless transcription, turn-taking, and quick responsiveness without the latency and fragility introduced by multiple providers and endpointing logic. Deepgram’s models are designed to enhance performance in complex environments like contact centers, healthcare, and financial services by maintaining accuracy and clarity in transcription and synthesis. The platform supports enterprise requirements with features like zero data retention, SOC 2 Type II compliance, and global data residency options, while the Together AI infrastructure offers high reliability and operational efficiency with its dedicated GPU capacity and unified SDKs.

Wan 2.7, a four-model suite launched on Together AI's platform, enhances video generation, continuation, and editing capabilities, initially focusing on text-to-video and soon expanding to image-to-video, reference-to-video, and video editing. It offers tighter creative control with features like optional audio, frame-level conditioning, and reference inputs, aiming to reduce workflow fragmentation by keeping all processes within a single platform rather than disparate tools. The Wan 2.7 suite, available at $0.10 per second of generated video, includes flexible resolution outputs, duration control, and prompt-driven direction, and will soon introduce more advanced workflows for image-to-video and reference-to-video models. The video editing model will allow for instruction and reference editing, style transfer, and temporal feature transfer, ensuring a cohesive workflow without the need to switch between different systems, thereby streamlining production from development to enterprise deployment with consistent APIs and billing.

A recent study explores how large language models (LLMs) can enhance database query optimization by addressing a key limitation in traditional query optimizers: their inability to understand semantic correlations in data. Through the introduction of DBPLANBENCH, a system that interacts with the Apache DataFusion engine, the study demonstrates significant improvements in query execution times and resource usage without altering the underlying database engine. This is achieved by translating complex physical execution plans into more manageable representations and applying targeted edits via evolutionary search. The system's efficiency is highlighted through case studies showing speedups of up to 4.78x in certain multi-join query scenarios, proving that LLMs can effectively function as semantic cardinality estimators. Moreover, the study establishes a practical workflow where optimizations found using smaller scale databases can be successfully transferred to larger ones, thereby validating the "optimize small, deploy large" approach. These findings indicate that AI and LLMs can not only support but also actively optimize the functional components of large-scale systems.

Aurora is an open-source, reinforcement learning-based framework designed to address the limitations of speculative decoding in production environments. It continuously learns and updates from live inference traces, unlike traditional static speculators that often become stale and ineffective as traffic patterns shift. Aurora's design allows it to adapt in real-time across various domains, offering a 1.25x speedup over well-trained static speculators and reducing infrastructure costs by eliminating the need for large-scale offline activation-collection pipelines. The framework supports diverse user demands and is algorithm-agnostic, making it compatible with future speculator designs. Aurora's serve-to-train flywheel approach, powered by RL, ensures efficient, non-disruptive updates, aligning training with real deployment utility rather than just offline quality. Through experiments, Aurora has demonstrated robust online adaptation and performance improvements, challenging the conventional reliance on extensive offline pretraining for speculative decoding.

On Memorial Day 2022, Dan Fu, Tri Dao, and their colleagues challenged the AI establishment by publishing FlashAttention, which demonstrated significant GPU performance improvements by focusing on memory movement and compute patterns. This breakthrough not only garnered attention from AI leaders like Andrej Karpathy but also highlighted the overlooked potential in GPU optimization. Their work emphasized the crucial role of efficient software, or kernels, in bridging the gap between AI models and hardware capabilities, a concept further advanced by their ThunderKittens library, which dramatically simplifies code for new GPU generations. The efforts of Together AI, with their academic-industry partnership model, have led to substantial performance gains in AI applications, as seen in their Megakernel project, which significantly reduced latency for a real-time voice agent company. This approach underscores the necessity of optimized AI infrastructure for the AI Native Cloud, where custom solutions tailored to specific workloads can make a decisive impact on performance and scalability.