MCP App StoreLe meilleur de l'écosystème IA et MCP, sélectionné chaque jour.
Sources
Hugging Face releases an open-source implementation of deep research agents capable of complex search and synthesis. This provides a transparent alternative to closed research agents and advances open-source agentic workflows.
DABStep introduces a new benchmark specifically designed for data agents performing multi-step reasoning. This provides a more rigorous way to evaluate how agents handle complex data manipulation and analysis tasks.
The first update on Open-R1, an open-source effort to replicate and improve upon DeepSeek-R1's reasoning capabilities. Focuses on making high-performance reasoning models accessible and transparent for the community.
A showcase of the DeepSeek-R1-Distill-Qwen-1.5B model's reasoning capabilities using a countdown game. Demonstrates efficient small-model distillation for complex logical tasks.
Guide on deploying and fine-tuning DeepSeek R1 models using AWS infrastructure. Essential for developers looking to optimize high-performance reasoning models in a scalable cloud environment.
Hugging Face introduces Inference Providers on the Hub, streamlining how developers access and deploy models via third-party providers. This simplifies the transition from model discovery to production API usage.
Hugging Face releases Open-R1, a fully open-source reproduction of the DeepSeek-R1 model. This effort democratizes advanced reasoning capabilities and provides a transparent baseline for open AI research.
Hugging Face has integrated Vision Language Models (VLMs) into smolagents, enabling AI agents to process visual information. This expansion allows for more complex agentic workflows that combine reasoning with visual perception.
NVIDIA introduces KVPress, a technique designed to optimize the management of Key-Value (KV) caches in LLMs to handle longer contexts more efficiently. This improvement reduces memory overhead and enhances the performance of models dealing with massive input sequences.
Hugging Face introduces 256M and 500M parameter versions of SmolVLM, significantly reducing the footprint of vision-language models. This enables more efficient local deployment and faster inference for edge AI applications.
Hugging Face partners with FriendliAI to streamline the deployment of models from the Hub. The integration aims to reduce the friction between model discovery and production-ready hosting for developers.
Hugging Face integrates the timm library into the transformers ecosystem, allowing developers to use a vast array of computer vision models seamlessly. This expands the accessibility of state-of-the-art vision architectures.
TGI now supports multiple backends, including NVIDIA TRT-LLM and vLLM, offering greater flexibility and performance optimization for LLM serving. This allows developers to choose the best execution engine for their specific hardware and model.
Hugging Face introduces a method to train static embedding models up to 400x faster using Sentence Transformers. This significantly lowers the barrier for creating efficient, specialized embedding models for RAG and search.
Hugging Face expands Visual Document Retrieval (VDR) capabilities to support multiple languages. This enables more effective cross-lingual document search and analysis for developers working with global datasets.
Hugging Face introduces smolagents, a lightweight library where agents perform actions by writing actual Python code. This approach simplifies agent logic and improves reliability compared to traditional JSON-based tool calling.
A technical deep dive into visualizing and managing GPU memory within PyTorch to optimize model training. Essential for developers dealing with OOM errors and memory fragmentation.
Explore NVIDIA's LogitsProcessorZoo for advanced control over LLM generation. This tool provides a systematic way to constrain and steer model outputs, essential for developers building reliable agentic workflows.
Hugging Face introduces Big Bench Audio to benchmark audio reasoning capabilities in AI models. This release provides a standardized framework for evaluating how models interpret and reason across complex audio datasets.
Hugging Face introduces ModernBERT, a modernized encoder model designed to replace BERT with significantly better performance and efficiency. It brings current LLM architectural improvements to the encoder space, benefiting a wide range of NLP tasks.