MCP App StoreThe latest from the AI and MCP ecosystem, curated daily.
Sources
Hugging Face introduces OpenEnv, an open-source initiative to democratize Agentic Reinforcement Learning. The project focuses on creating standardized environments for agent training and evaluation.
Exploration of a multi-agent economy simulated using a small 3B parameter model. Demonstrates the viability of complex agent interactions and emergent behavior in highly constrained model sizes.
NVIDIA releases Nemotron 3.5 Content Safety, providing customizable multimodal safety guardrails for enterprise AI. It focuses on scaling safety checks across global deployments and various data modalities.
ServiceNow AI introduces EVA-Bench Data 2.0, a comprehensive dataset for evaluating agentic AI across 121 tools and 213 scenarios. This provides a robust framework for measuring tool-use and task-completion capabilities of AI agents.
NVIDIA introduces a new method for generating high-quality synthetic Q&A pairs to improve the pretraining of Nemotron models. This approach leverages task-seeded generation to create more diverse and accurate training data.
Hugging Face introduces a redesigned CLI optimized for AI agents to interact with the Hub. This enables more seamless model and dataset management within agentic workflows.
Dharma-AI explores the application of Direct Preference Optimization (DPO) outside of traditional chatbot contexts. This research expands the utility of preference alignment for broader AI tasks.
A practical implementation of the Model Context Protocol (MCP) to add tools to the Reachy Mini robot. This showcases how MCP can bridge LLMs with physical robotic hardware.
Hcompany introduces Holo3.1, focusing on high-performance local computer use agents. This release emphasizes reduced latency and improved local execution for agentic UI interactions.
JetBrains has released Mellum2, a 12B parameter Mixture-of-Experts (MoE) model. The model is designed for high-performance AI tasks, continuing JetBrains' push into specialized developer-centric AI models.
IBM Research explores the necessity of 'agent logic' for scaling AI in enterprises. The piece argues that moving beyond simple LLM prompts to structured agentic workflows is key to reliable AI adoption.
NVIDIA has released Cosmos 3, an open omni-model designed for physical AI reasoning and action. This model aims to bridge the gap between digital intelligence and physical world interaction, providing a foundation for robotics and autonomous systems.
A comprehensive beginner's guide to using torch.profiler for PyTorch performance analysis. Essential for developers looking to identify bottlenecks and optimize model execution speed.
A new benchmark, ITBench-AA, reveals that even frontier models struggle with enterprise IT agentic tasks, with most scoring below 50%. This highlights a significant performance gap in deploying reliable AI agents for complex IT infrastructure management.
Hugging Face introduces Delta Weight Sync in TRL, enabling efficient distribution and syncing of massive model weights (up to a trillion parameters) using Hub Buckets.
Hugging Face establishes a clear taxonomy for AI agent concepts, defining critical terms like 'harness' and 'scaffold'. This glossary provides developers with a standardized language to describe agent architectures and their operational environments.
NVIDIA introduces diffusion-based language models from Nemotron-Labs aimed at achieving near-instantaneous text generation. This research explores a fundamental shift in how LLMs generate tokens to drastically reduce latency.
Explores the shift from massive general-purpose models to smaller, specialized ones. Discusses how targeted training and architecture can outperform larger scales for specific developer tasks.
AllenAI releases OlmoEarth v1.1, a more efficient family of models designed for geospatial and earth-science applications. This improves accessibility and performance for specialized AI research in environmental domains.
Hugging Face introduces the Ettin Reranker family, designed to improve the precision of retrieval-augmented generation (RAG) pipelines. A key tool for developers looking to refine document ranking in AI applications.