MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
Sources
Hugging Face introduces SmolLM3, a compact multilingual model designed for efficient reasoning with long-context support. Ideal for on-device deployment and lightweight agentic tasks.
Hugging Face introduces a streamlined MultiModal Data Pipeline (MMDP) designed to optimize the handling and processing of diverse data types for AI training. The pipeline focuses on efficiency and scalability for multimodal model development.
A technical guide on implementing sparse embedding models using Sentence Transformers. This is highly relevant for developers building advanced RAG systems and information retrieval pipelines.
NVIDIA has released the Llama Nemotron Nano VLM on Hugging Face, providing a compact vision-language model for efficient edge deployment and specialized AI applications.
Anthropic launched Desktop Extensions — a one-click installation system for MCP servers directly in Claude Desktop, eliminating the need to manually edit JSON config files. Users can browse, install, and manage MCP servers from a UI without touching the terminal. A major UX improvement that significantly lowers the barrier to MCP adoption.
Google's Gemma 3n is now fully available in the open-source ecosystem. This release continues the push for democratizing high-performance AI models for the developer community.
SGLang now integrates a Transformers backend, enhancing the flexibility and compatibility of the serving framework. This allows developers to more easily deploy and optimize a wider range of models.
Anthropic walks through the engineering decisions behind Claude's Research feature — a multi-agent system that parallelises web search, synthesises findings across agents, and produces long-form research reports. The post covers orchestration architecture, reliability challenges, and how they handled conflicting information between sub-agents. Valuable real-world reference for anyone building research or RAG-style multi-agent systems.
Analysis of how long prompts can block concurrent LLM requests and degrade system performance. Essential reading for developers optimizing throughput and latency in production AI environments.
Introduction to the Hugging Face Kernel Hub, providing a fast track for developers to explore and deploy kernels. A useful resource for those looking to extend HF's computational capabilities.
Hugging Face integrates Featherless AI as an inference provider, expanding options for deploying and serving open-source models. This move simplifies access to high-performance inference for developers building with the HF ecosystem.
Hugging Face and NVIDIA have launched Training Cluster as a Service, simplifying the deployment of massive compute resources for model training. This collaboration lowers the barrier for developers to scale their training infrastructure efficiently.
ScreenSuite is a new comprehensive evaluation framework specifically designed for GUI agents. It provides a robust set of benchmarks to measure how effectively AI agents can navigate and interact with graphical user interfaces.
A technical deep-dive into implementing KV Caching within nanoVLM. This guide provides a foundational look at optimizing inference efficiency for small language models.
HCompany introduces Holo1, a new family of Vision-Language-Action Models (VLMs) designed for GUI automation. These models power the Surfer-H agent, improving the ability of AI to interact with and navigate complex graphical user interfaces.
Hugging Face releases SmolVLA, an efficient Vision-Language-Action model trained using LeRobot community data. This model aims to democratize robotic control by providing a lightweight, open-source VLA for embodied AI tasks.
Hugging Face introduces co-located vLLM in TRL to maximize GPU efficiency during reinforcement learning. This optimization allows for better resource utilization when training agents with complex feedback loops.
A new approach to CodeAgents that emphasizes structured action execution for better reliability. By combining code generation with formal structures, agents can more accurately perform complex tool-use tasks.
Liger GRPO is now integrated with TRL, enabling more efficient Group Relative Policy Optimization for LLM training. This allows developers to scale reinforcement learning from human feedback with significantly reduced memory overhead.
A practical guide demonstrating how to build a functional AI agent using the Model Context Protocol (MCP) in just 70 lines of Python. This highlights the accessibility and efficiency of the MCP ecosystem for developer tooling.