July 2026 Summaries

Moonshot AI's Kimi K3, released on July 16, 2026, represents a significant advancement in open-weight AI models with its 2.8-trillion-parameter reasoning capability and multimodal input support. Utilizing a Mixture of Experts architecture with Stable LatentMoE and innovations like Kimi Delta Attention and Attention Residuals, the model offers enhanced scalability and reasoning efficiency. Kimi K3 excels in complex, long-horizon tasks, achieving a high Intelligence Index score well above the median for open-weight models. Despite its premium pricing and relatively high latency, it is positioned as a strong choice for developers needing extensive reasoning capabilities, with multiple API providers offering varied pricing and performance options. Fireworks, Together AI, and OpenRouter present diverse benefits from cost-efficient training to reliable production infrastructure, while Moonshot's first-party API provides a reference implementation for benchmark alignment. The model's extensive context window and open weights make it particularly suitable for coding, research, and enterprise workflows, although its slower output speed suggests it is best used for asynchronous tasks prioritizing accuracy over speed.

DeepInfra's analysis of AI models Kimi K3, Claude Opus 4.8, and GPT-5.6 Sol highlights their competitive capabilities in context window size, vision support, and pricing, with all models scoring closely on the Artificial Analysis Intelligence Index. The comparison focuses on structural differences such as cost efficiency, particularly for workloads involving long-context and agentic tasks, where K3 emerges as the most cost-effective option despite the absence of a non-thinking mode. Opus 4.8 excels in mathematics and GUI agent tasks, while Sol leads in software engineering benchmarks like Terminal-Bench. Additionally, K3 is the only model with open weights available via Hugging Face, making it a viable option for teams requiring on-premise deployment and data sovereignty. The decision-making process is influenced by context length distribution and the need for open weights, with transparency and documentation gaps noted across the models.

In 2026, three Chinese AI labs released significant open-weight AI models—Kimi K3, DeepSeek V4 Pro, and GLM-5.2—that challenge the notion of open-weight models being budget alternatives to closed systems. Kimi K3, developed by Moonshot AI, is the largest with 2.8 trillion parameters and features native multimodal capabilities, placing fourth on the Artificial Analysis Intelligence Index. DeepSeek V4 Pro excels in cost efficiency and competitive programming, holding a top position on LiveCodeBench, while GLM-5.2 offers superior throughput and focuses on latency-sensitive applications. Each model is licensed under MIT or Modified MIT terms and presents distinct capabilities, benchmarks, and architecture nuances, with implications for deployment logistics and production choices.

DeepInfra has introduced Hosted Agents, a service that simplifies the deployment of AI agents by providing a fully managed infrastructure, starting at $13 per month. The service offers two types of agents: OpenClaw, which includes a web dashboard for team use and integration with communication tools like Slack and WhatsApp, and Hermes, a command-line driven option for individual developers. Both types offer two hosting tiers to cater to different workloads and are designed to be inference-ready from the start, eliminating common setup hurdles. The platform ensures seamless updates, automatic backups, and long-term memory retention, providing an isolated and private environment for each agent. Users benefit from cost-effective pricing, with no fees for idle agents and separate billing for inference based on usage. These features make it an attractive solution for developers and teams seeking reliable AI assistants without the burden of maintaining underlying systems.

DeepInfra conducted independent benchmarking of NVIDIA's new Vera CPU, designed for agentic AI applications, to evaluate its performance in handling AI agent workloads. The testing revealed that Vera significantly outperforms other leading CPUs, such as Intel's Granite Rapids and AMD's Zen5, demonstrating a 2.2× increase in orchestration-step latency and the ability to sustain more concurrent agents at a given quality of service. This performance is attributed to Vera's efficient per-core execution and ability to handle complex, branch-heavy workloads with minimal cache misses. The CPU also demonstrated strong capabilities in handling additional tasks, such as inference, even when fully loaded with agent microVMs, and showed potential for reducing overhead costs associated with secure-isolation layers. These findings support NVIDIA's claim that traditional infrastructure designed for previous architectures may not be optimal for agent workloads, underscoring the importance of workload-specific hardware design.

DeepInfra has announced the integration of NVIDIA Nemotron 3 Embed models into its platform, offering advanced retrieval capabilities for agents that require accurate and cost-effective data processing. The Nemotron 3 Embed comes in two variants: the 8B model, which offers top-tier retrieval accuracy, and the 1B model, which retains over 95% of the larger model's accuracy while significantly reducing costs. Both models are optimized for NVIDIA Blackwell GPUs, enhancing throughput and efficiency. This development is crucial for multi-turn agents involved in complex tasks like planning, code understanding, and enterprise search, as strong retrieval minimizes irrelevant context and reduces the risk of errors. DeepInfra facilitates easy integration and experimentation with these models through an OpenAI-compatible API, allowing users to seamlessly transition from testing to production.

DeepInfra has introduced a new Flex service tier designed to offer users a cost-effective alternative for non-urgent computational tasks, providing a 20% discount compared to real-time pricing. This tier is intended for workloads that prioritize cost over immediate response, such as model evaluations, data enrichment, and asynchronous tasks, by allowing a wait time of up to 10 minutes for available capacity. The Flex service operates through the same API as standard requests, making it easy for users to switch by simply adding "service_tier": "flex" to their requests. If the request waits without being processed, no charges are incurred, and this new feature is compatible with any OpenAI-compatible models that support the tier, offering a seamless transition for existing users.

LangChain Deep Agents have achieved frontier-level performance on open models by optimizing the integration of NVIDIA Nemotron 3 Ultra with DeepInfra, significantly reducing costs compared to closed models. This integration emphasizes the importance of the "harness," which includes prompts, tools, and middleware that enhance the model's performance without the need for retraining. The result is top-tier agent accuracy at a fraction of the cost, making it accessible for developers to utilize in production via DeepInfra's platform. By leveraging the open model and harness, users can customize and deploy high-efficiency agents with ease, supported by DeepInfra's competitive pricing and infrastructure.

In mid-2026, DeepInfra announced OpenCode, an open-source alternative to Anthropic's Claude Code, designed to provide coding teams with flexible and cost-effective options for utilizing various AI models. The release came after the U.S. government restricted access to Anthropic's Fable 5, highlighting the need for a model-agnostic agent like OpenCode, which allows users to choose from over 75 providers, including open-weight models. OpenCode's design emphasizes model freedom, enabling users to select models based on budget and capability, unlike Claude Code, which is optimized for Anthropic's ecosystem. OpenCode features a modular architecture with built-in Git-based snapshots and Language Server Protocol support for real-time error correction, providing a robust alternative that adapts to different coding needs and budgets. The discussion emphasizes that for OpenCode to be a viable choice, the underlying models must be both cost-effective and capable, with models like Zhipu's GLM 5.2 offering strong performance at competitive prices. This flexibility is particularly advantageous for teams seeking to avoid the risks of closed systems, such as sudden access revocations, while optimizing costs through strategic model selection and usage.

DeepSeek, a research-focused AI lab, is in the final stages of its first external financing round, aiming to raise approximately $10.29 billion, with a pre-money valuation of around $45 billion. The lab, which previously relied on internal funding from its parent hedge fund Zhejiang High-Flyer Asset Management, is seeking capital to support its ambitions in artificial general intelligence (AGI) and maintain its open-source commitment. The round involves contributions from founder Liang Wenfeng and potential investments from the National Artificial Intelligence Industry Investment Fund and other major firms like Tencent and JD.com. DeepSeek has gained significant attention for its cost-efficient, high-performance models, such as the V4 family, optimized for various silicon, challenging the notion that frontier AI development requires closed-source models. The financing round signifies a potential shift in the AI industry towards valuing open-source, research-driven approaches, but raises questions about maintaining open-source commitments amid rising training costs and the influence of strategic investors.

MiMo-V2.5, a Xiaomi model released in April 2026, stands out for integrating open weights, a 1 million-token model design, and competitive pricing, which varies depending on the provider. Notably, Xiaomi's first-party API offers the lowest rates, making it attractive for cost-sensitive developers, while DeepInfra provides managed deployment options with higher output prices, appealing to those seeking operational control and platform features. As an omnimodal system, MiMo-V2.5 supports text, image, video, and audio, and its practicality is underscored by an Intelligence Index score of 40 and an output speed surpassing comparable models. The model's adaptability to reasoning-style tasks and availability on platforms like DeepInfra and Hugging Face under the MIT license allows teams to choose between hosted and self-hosted solutions without changing model families. Ultimately, the selection of a provider should consider workload specifics, such as token reuse, output verbosity, and required platform features, rather than solely focusing on the lowest per-token cost.

DeepInfra, in collaboration with LangChain, offers a streamlined approach to building retrieval-augmented generation (RAG) applications, allowing users to keep the entire pipeline on a single OpenAI-compatible endpoint. This integration simplifies the process by combining document embedding and natural language generation under one account, thereby reducing the complexity of managing multiple API keys and billing systems. Users can leverage DeepInfra's advanced models like Qwen3-Embedding-8B for multilingual embeddings and DeepSeek-V3.2 for generation, which provides large-model output quality at a fraction of the cost. The process involves an offline indexing phase where documents are converted into searchable vectors, and a live retrieval and generation phase that uses these vectors to generate contextually grounded answers. The unified system also enhances operational efficiency as it adheres to a zero-retention policy, ensuring that both document text and queries are not stored for training purposes, thus simplifying the auditing process. This approach not only improves accuracy and efficiency but also offers a cost-effective solution for deploying scalable and robust RAG applications.

MiMo-V2.5, released by Xiaomi on April 22, 2026, is a highly advanced multimodal reasoning model optimized for complex reasoning and long-horizon tasks, featuring a sparse Mixture of Experts (MoE) architecture with 310 billion total parameters. The model's extensive 1 million-token context window and its ability to process text, image, video, and audio inputs make it particularly suitable for retrieval-augmented generation (RAG) and long-context document analysis. DeepInfra, a serverless inference platform, is recommended for its superior speed and low latency, achieving over 130 tokens per second, making it ideal for latency-sensitive applications. Conversely, Xiaomi's first-party API offers significant cost-efficiency and prompt caching benefits, with pricing as low as $0.14 per million input tokens and $0.28 per million output tokens, appealing to cost-sensitive batch workloads. MiMo-V2.5 is open-source and released under the MIT license, allowing for unrestricted commercial use, with all weights available on Hugging Face.

DeepInfra's analysis of the GLM-5.2 (max) Mixture-of-Experts model, with its substantial 753 billion parameters and a 1 million token context window, highlights the model's demanding infrastructure requirements for optimal deployment. The guide evaluates top API providers based on throughput, latency, pricing, and quantization architecture, naming DeepInfra as the overall recommended provider due to its use of FP4 quantization, which strikes a balance between cost-efficiency and performance. Fireworks emerges as the leader in raw speed with a throughput of 314.9 tokens per second, while GMI offers the most cost-effective solution at $0.72 per 1 million tokens using FP8 quantization. The report underscores the importance of selecting providers capable of handling the model's extensive reasoning capabilities and context window, considering factors like throughput, latency, and cost to ensure efficient deployment of GLM-5.2 (max).

Xiaomi's MiMo-V2.5 is a groundbreaking model that unifies agentic and multimodal capabilities, previously managed by two separate models, into one efficient architecture. It processes text, images, video, and audio, extending its context to 1 million tokens and surpassing its predecessors in agentic and multimodal benchmarks, showcasing significant improvements in tasks like video understanding. The model's architecture, featuring a sparse Mixture-of-Experts design with 310 billion parameters and only 15 billion active per forward pass, ensures economic practicality. A hybrid attention mechanism underpins its 1M-token context window, optimizing performance without sacrificing throughput. Open-sourced under the MIT license, MiMo-V2.5 is available on DeepInfra, offering an OpenAI-compatible API with flexible deployment options and straightforward pricing. The model's consolidation of diverse input handling and agentic task performance into a singular efficient framework positions it as a valuable tool for applications requiring comprehensive perception, reasoning, and action capabilities.

GLM-5.2, developed by Z-AI and available on DeepInfra, is a cutting-edge model characterized by a stable 1,048,576-token context window, which is crucial for long-horizon tasks. This advancement, achieved through the new IndexShare architecture, significantly reduces computational costs by reusing indexers across sparse attention layers. The model, which also includes an enhanced multi-token prediction layer for better speculative decoding, is licensed under MIT with no regional restrictions, allowing developers greater flexibility in deployment. It excels in coding tasks, outperforming its predecessor GLM-5.1 with notable benchmark improvements, and supports both English and Chinese natively. GLM-5.2 is accessible via an OpenAI-compatible API with tiered pricing based on usage, and it maintains a zero data-retention policy, ensuring data privacy and security.

By mid-2026, the gap between open-source and closed-source AI models has significantly narrowed, with open models now competing effectively in areas like mathematics, general knowledge, and graduate-level science reasoning. The release of open models like DeepSeek and Qwen has contributed to this convergence, offering performance on par with or surpassing proprietary models in many benchmarks. While closed models still hold an advantage in complex tasks requiring nuanced human interaction, safety calibration, and the latest multimodal capabilities, open models are increasingly becoming the economically rational choice for most production workloads due to their cost-effectiveness and versatility. The shift in the AI landscape is further underscored by the rise of Chinese open-source models, which now dominate global download rankings, reflecting a broader geographic reorientation in AI development. For most applications, the decision between open and closed models is now more about cost and control rather than a clear quality difference.

DeepInfra has published a guide on the best Software as a Service (SaaS) tools and API providers for utilizing the MiMo-V2.5 model series, which is notable for its advanced multimodal capabilities and large context handling. As the complexity of large language models (LLM) increases, MiMo-V2.5 introduces a 1 million-token context window and supports text, image, video, and audio processing, requiring significant infrastructure considerations. The guide highlights various service providers, including DeepInfra, which offers a scalable and cost-effective API solution, and Xiaomi, which provides direct access and low-latency performance through their Token Plan. Other providers like OpenRouter, Kilo Code, TypingMind Teams, The Grid, and LMSpeed offer unique features such as multi-model routing, IDE integration, team-oriented UI workspaces, spot-pricing for cost reductions, and API speed testing. The guide suggests that DeepInfra stands out as the top choice for enterprises seeking a robust, scalable, and budget-friendly foundation for deploying MiMo-V2.5, while other options cater to specific needs such as direct model access, rapid deployment, and budget optimization.

AI subscription fatigue arises from the cumulative cost and complexity of managing multiple AI tools, each with separate subscriptions, interfaces, and integration requirements. This issue parallels the subscription fatigue faced by consumers in other areas, like streaming services, but is exacerbated in the AI sector due to the need for tools to work together seamlessly. The proposed solution is to consolidate AI subscriptions into a singular, pay-as-you-go API model, exemplified by DeepInfra's offering. This approach would eliminate the fragmentation and overhead of managing multiple subscriptions, offering a unified billing system based on token consumption rather than seat-based or flat-tier pricing. This model allows users to access a wide range of AI models from a single account, optimizing cost and efficiency by matching each task with the most appropriate and cost-effective model in the catalog, thereby addressing the root cause of subscription fatigue rather than just its symptoms.

DeepInfra's GLM-5.2 is a cutting-edge large language model developed by Z.AI for complex reasoning, software engineering, and extensive data processing tasks, featuring an unprecedented 1,048,576-token context window and significant architectural enhancements. Released on June 13, 2026, it introduces innovations like IndexShare for efficient context window usage and an upgraded Multi-Token Prediction layer to improve decoding speed and cost-effectiveness. With its flexible reasoning system and high performance on industry-standard benchmarks, GLM-5.2 rivals proprietary models such as GPT-5.5 and Claude Opus 4.8, showcasing notable achievements in mathematical excellence and agentic orchestration. The model is accessible via DeepInfra’s OpenAI-compatible API, facilitating straightforward integration for developers, and is available under the MIT license for unrestricted commercial use. Pricing is based on a flexible, pay-per-token model, with options for standard and prioritized workloads, and users can deploy Private Endpoints for dedicated capacity.

XiaomiMiMo's MiMo-V2.5 is an advanced omnimodal AI model designed to process and understand diverse data types such as text, image, video, and audio through a unified architecture leveraging a 310-billion-parameter Sparse Mixture of Experts framework, activating only 15 billion parameters during inference. This model offers a substantial context window of 1 million tokens, supporting complex multimodal perception and autonomous workflows, while integrating native encoders for various data types to enhance cohesion. MiMo-V2.5 showcases significant improvements over its predecessor, particularly in reasoning and computational efficiency, by employing a hybrid attention architecture and multi-token prediction modules that enhance inference speed and reinforcement learning efficacy. Hosted on DeepInfra, it provides high-performance, low-latency inference via an API compatible with OpenAI, making it a versatile choice for developers aiming to implement agentic workflows and process extensive document sets. Pricing is usage-based, with options for standard and priority tiers to optimize cost and processing speed, making the model accessible for professional-grade deployments.

DeepInfra's comparison of three open-weight models, DeepSeek V4 Flash, Qwen3.6 35B A3B, and GLM-4.6, highlights their distinct architectural features and intended use cases on the inference cloud platform. DeepSeek V4 Flash is optimized for cost-effective reasoning with a large parameter capacity but only activates a small portion per token, making it ideal for long-document analysis on a budget. Qwen3.6 35B A3B excels in agentic coding and multimodal inputs, balancing speed and efficiency with the ability to process text, images, and videos. GLM-4.6, designed for tool use and long-context retrieval, offers a 200k token window and targets workflows requiring intricate tool interactions. Each model's performance is assessed based on intelligence, speed, and inference cost, with DeepSeek being the most cost-efficient, Qwen the fastest, and GLM offering advanced retrieval capabilities at a higher price point. The choice among these models depends on specific requirements, such as cost sensitivity, multimodal input processing, or tool-heavy agent loops.

GLM-5.2 is a groundbreaking open-weight model designed to handle complex reasoning, long-context processing, and agentic coding tasks with its expansive 1-million token context window and Mixture-of-Experts (MoE) architecture. DeepInfra, a top provider, offers a balance of cost and low latency, making it suitable for real-time applications and Retrieval-Augmented Generation (RAG) pipelines with a competitive price of $0.80 per 1 million tokens. Other providers like Fireworks AI and Z.ai cater to speed and direct ecosystem access, respectively, while Scaleway and Gleap focus on European data sovereignty and secure EU-based infrastructure. The guide highlights the best SaaS tools and API providers for deploying GLM-5.2, emphasizing the importance of performance benchmarks, pricing, and enterprise requirements, and showcases how different providers excel in specific areas such as throughput, scalability, and compliance with strict data privacy laws.

GLM 5.2, a reasoning model from Z.ai, stands out for its 1M-token context window and open weights, offering flexibility in deployment and cost tuning for teams seeking long-context reasoning without being locked into a closed model. Released on June 16, 2026, it features 753 billion total parameters, and its performance is considered strong, with an Intelligence Index score of 51, placing it above many comparable models. Available through hosted APIs and public model weights, GLM 5.2 is commercially usable under the MIT license, with pricing that varies among 25 providers, ranging from $0.95 to $3.00 per million input tokens and $3.00 to $10.25 per million output tokens. Despite being relatively expensive compared to its peers, its ability to support long-running workflows, project-scale coding, and document-heavy pipelines makes it a practical choice for developers and machine learning teams. DeepInfra emerges as a favorable platform for hosting GLM 5.2 due to its competitive pricing and infrastructure, offering both public API access and private deployment options for production use.