MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
Sources
H company released Holotron-12B — a 12B parameter model optimised for high-throughput computer use, balancing strong automation performance with the speed needed for real-time desktop and browser control. Positioned as a cost-effective option for teams that need computer use at scale without paying frontier model prices.
Hugging Face launched Storage Buckets — a new Hub feature for storing arbitrary files (datasets, checkpoints, outputs) alongside models and datasets in a unified workspace. Removes the need to use separate S3/GCS buckets for ML artefacts, keeping everything in the HF ecosystem. A practical platform improvement for teams with large training pipelines.
A comparative analysis of 16 open-source reinforcement learning libraries for LLM training — covering throughput, async rollout support, scalability, and ease of use. Cuts through the rapidly crowded RL-for-LLMs space with concrete benchmarks and recommendations for different training scales. Useful reference for teams choosing a post-training RL stack.
GGML and llama.cpp — the foundational libraries behind local LLM inference — are joining Hugging Face to ensure their long-term development and maintenance. This brings the most widely-used local inference stack under HF's open-source umbrella, with commitments to continued compatibility and community governance. A significant consolidation in the local AI ecosystem.
IBM Research and UC Berkeley released IT-Bench and MAST — a benchmark suite and failure taxonomy for enterprise AI agents in real IT environments. The research identifies the root causes of enterprise agent failures across categories like tool misuse, context loss, and ambiguity handling. Practical reference for teams debugging why their production agents underperform in enterprise settings.
Hugging Face demonstrates using Codex and Claude agent skills to generate custom CUDA kernels — making GPU-level optimisation accessible to developers without deep CUDA expertise. The post shows how agent skills can encode specialist knowledge (GPU programming patterns) and apply it on demand. A compelling use case for agent skills beyond typical software engineering tasks.
OpenEnv is a new evaluation framework that tests tool-using agents in real-world execution environments rather than sandboxed simulations — catching failure modes that controlled benchmarks miss. The post covers the framework design, evaluation methodology, and results across several agent models. Useful for any team building agents that interact with real APIs or filesystems.
Transformers.js v4 is now available on NPM, bringing Hugging Face's in-browser and Node.js inference library to a major new release with improved model support, better performance, and a cleaner API. Developers can now run transformer models directly in JavaScript without a Python backend. A big upgrade for web developers building AI-powered browser applications.
Hugging Face launched Community Evals — an open framework for community-run model evaluations that anyone can contribute to, challenging the opacity of existing leaderboards. The goal is to build a more trustworthy, diverse set of model benchmarks driven by real use cases rather than optimised for leaderboard gaming. Important infrastructure for the open-source AI evaluation ecosystem.
Hugging Face reflects on one year since DeepSeek disrupted the AI landscape, examining how open-source AI has evolved in response — from the proliferation of capable open-weight models to new training approaches and the emerging AI+ ecosystem. The third in a series, focusing on where the global open-source AI movement goes next.
LinkedIn's AI team shares a practical retrospective on training GPT-OSS with agentic reinforcement learning — covering the infrastructure challenges, reward design decisions, and lessons learned from scaling RL training to production-grade agent models. One of the more candid engineering accounts of what it actually takes to do agentic RL at scale.
Hugging Face marks one year since DeepSeek-R1 shocked the AI world by matching frontier performance at a fraction of the cost, fundamentally changing assumptions about what open-source models can achieve. The post reflects on how that moment catalysed a wave of open-weight model development and reshaped the competitive landscape between open and closed AI.
Hugging Face launched Open Responses — an open-source implementation of the Responses API spec, letting developers use the same agent-friendly API pattern against any open-weight model instead of being locked to OpenAI's hosted version. Compatible with OpenAI's SDK, making migration straightforward. Important infrastructure for the open-source agent ecosystem.
ServiceNow AI presents AprielGuard, a new safety guardrail designed to improve adversarial robustness in LLM systems. The framework focuses on preventing malicious prompts and ensuring safe model outputs in production environments.
IBM Research introduces CUGA on Hugging Face, a framework for creating and sharing configurable AI agents. This initiative aims to democratize agent development by providing a standardized way to define and deploy agentic behaviors.
llama.cpp introduces new model management capabilities, improving how local LLMs are handled. This is a key update for developers building edge-AI applications and local inference pipelines.
Codex is open-sourcing its AI models to promote transparency and democratize access to high-quality code generation tools. This move supports the open-science movement in AI development.
Hugging Face releases swift-huggingface, a comprehensive Swift client for integrating HF models and datasets into Apple ecosystem apps. This enables native AI integration for iOS and macOS developers.
Intel introduces DeepMath, a lightweight math-reasoning agent built using the smolagents framework. It demonstrates efficient agentic workflows for solving complex mathematical problems with minimal overhead.
Hugging Face demonstrates using Claude to fine-tune open-source LLMs, automating the process of creating high-quality synthetic data for training. This highlights the power of frontier models in improving the open-source ecosystem.