Hugging Face introduces MolmoMotion, a language-guided 3D motion forecasting model. The system combines vision and language to predict future trajectories from videos, enabling applications in robotics and animation.
Hugging Face and Strands integrate Hub models with LeRobot to deploy AI agents on robot hardware. The platform enables developers to use pre-trained models to control physical robots directly.
Hugging Face announces GLM-5.2, a model designed for long-horizon tasks. The model improves capacity to handle extended contexts and complex multi-step workflows.
Hugging Face introduces agentic resource discovery, enabling AI agents to autonomously search and access models, datasets, and tools available on the platform. This capability enhances agent autonomy in executing complex tasks.
Hugging Face releases olmo-eval, an evaluation workbench for the model development loop. The tool automates performance testing and enables rapid iteration during language model training and fine-tuning.
Part 2 of a PyTorch profiling guide. Detailed performance analysis of nn.Linear layers and construction of an optimized fused MLP. Demonstrates operation fusion techniques to reduce latency and improve computational efficiency.
Hugging Face benchmarks frontier ASR models (Whisper, Canary, Conformer) on code-switched bilingual speech. Results reveal significant performance gaps across language pairs and models, exposing limitations of current voice agents for multilingual customer service.
Cohere releases North Mini Code, its first model designed for developers. The model is optimized for code generation and completion, with multilingual support and integration into the Hugging Face ecosystem.
An AI agent built a 3D Paris gallery by chaining two Hugging Face Spaces. The system automatically orchestrated image generation and 3D environment creation without manual intervention.
A researcher fine-tuned a language model to generate content optimized for ADHD brains by maximizing dopaminergic engagement. The approach combines fine-tuning on curated data and evaluation via behavioral metrics.
Hugging Face releases a migration guide for moving GitHub CI/CD pipelines to Hugging Face Jobs. The platform provides a native alternative to automate testing and deployments within the Hugging Face ecosystem.
Hugging Face study on emergent behaviors in a multi-model system. Simulation of a five-model AI economy showing how crashes can occur and vanish based on agent interactions. Analysis of control and emergence phenomena in complex systems.
Hugging Face built Pakistan Notice Helper, a small AI tool addressing a specific local safety problem in Pakistan. The tool leverages lightweight models tailored to regional needs and resource constraints.
The open source community is backing OpenEnv, a platform for agentic reinforcement learning. The project gains growing adoption and active collaboration within the open source ecosystem.
Hugging Face introduces Room360, a video-to-3D spatial reconstruction platform. The tool converts video sequences into usable 3D models for immersive and architectural applications.
OpenAI offers Codex vouchers to Hugging Face sponsors to test the code generation model. Partnership initiative between OpenAI and the community platform.
Hugging Face releases Her, a debugging tool for Claude Code sessions. Her analyzes interactions and identifies issues in AI-assisted coding workflows.
Hugging Face showcases a collaborative experiment across five labs using small models to build a dramatized financial scenario. The project demonstrates how reduced-size models can be orchestrated to generate complex narratives in a specialized domain.
Hugging Face launches Job Searcher, an AI-powered job search tool that helps candidates find relevant positions. The tool uses language models to analyze job postings and match candidate profiles.
Hugging Face introduces Persona Atlas, a tool mapping the thinking styles of famous personalities. The project analyzes cognitive patterns and decision-making approaches of public figures using language models.
Hugging Face deploys a multi-agent economy on a 3B (3 billion parameter) model. The 'Thousand Token Wood' system enables autonomous agents to interact, negotiate, and exchange resources in a simulated environment with constrained token budgets.
NVIDIA releases Nemotron 3.5 Content Safety, an open-source multimodal safety model detecting harmful content across text, image, and video. Customizable for global enterprises, it provides granular policy control for moderation across regions and use cases.
Hugging Face releases a guide to fine-tune Nemotron 3.5 ASR, NVIDIA's speech recognition model. The method enables adapting the model to specific languages, domains, or accents through fine-tuning.
Hugging Face releases EVA-Bench Data 2.0, a benchmark spanning 3 domains, 121 tools, and 213 scenarios to evaluate multi-tool AI agents. Major expansion from previous version to test models' ability to orchestrate complex tool interactions.
Nvidia releases a task-seeded synthetic Q&A generation method for Nemotron pretraining. The technique uses task seeds to create diverse, high-quality training data, improving model performance across varied benchmarks.
Hugging Face redesigns its CLI to optimize it as an agent. The command-line interface becomes agent-friendly with structured commands and parsable responses, enabling autonomous systems to interact directly with the Hub.
Hugging Face explores applying DPO (Direct Preference Optimization) beyond chatbots, including for vision and reasoning model optimization. The article details how this alignment technique can improve performance on complex tasks without requiring an explicit reward model.
Hugging Face integrates MCP (Model Context Protocol) tools into Reachy Mini, a humanoid robot. This integration enables the robot to access external tools via MCP protocol, expanding its interaction and autonomy capabilities.
Hugging Face releases Holo3.1, a fast local computer use agent for task automation. The model runs on-device without cloud dependency, enabling speed and privacy for system-level actions.
JetBrains releases Mellum2, a 12B Mixture-of-Experts model. The model combines computational efficiency with performance, designed for code and reasoning tasks.
Hugging Face argues that enterprise AI adoption beyond LLMs requires scalable agent logic. The article explores how multi-agent systems and orchestration become critical for deploying AI beyond simple use cases.
NVIDIA releases Cosmos 3, an open omni-model for physical AI that reasons and acts. The model processes video, text, and images to understand real-world physics and generate robotic actions.
Beginner's guide to PyTorch profiling using torch.profiler. Covers how to measure performance and identify bottlenecks in AI models, with practical examples for newcomers.
ITBench-AA, a new benchmark from Artificial Analysis and IBM, evaluates frontier models on agentic enterprise IT tasks. Top models (Claude, GPT-4, Gemini) score below 50%, exposing significant gaps in automating complex IT workflows.
Hugging Face introduces Delta Weight Sync in TRL to optimize deployment of trillion-parameter models. The technique syncs only weight changes rather than full models, drastically reducing storage and bandwidth requirements for updates.
Reachy Mini, Pollen Robotics' humanoid robot, now runs fully locally without cloud dependency. Integrates open-source models (Llama, Whisper) for vision, speech, and motor control. Deployed on embedded hardware.
Hugging Face clarifies AI agent terminology: distinguishing harness (execution infrastructure), scaffold (coordination structure), and agent (autonomous system). Essential definitions to avoid confusion in the ecosystem.
Nvidia and Hugging Face introduce Nemotron-Labs, diffusion-based language models to accelerate text generation. The approach parallelizes token generation, reducing latency compared to traditional autoregressive methods.
Hugging Face argues that AI model specialization outperforms raw scale in procurement decisions. Organizations typically favor large generalist models, overlooking that smaller specialized models deliver better performance and lower costs for specific tasks.
Hugging Face releases OlmoEarth v1.1, a more efficient family of models for geospatial tasks. The new models deliver improved performance and inference speed compared to the previous version.
Hugging Face introduces the Ettin Reranker family, models designed to improve search relevance and RAG result ranking. These rerankers optimize document ranking after initial retrieval.
Hugging Face releases a guide for fine-tuning NVIDIA Cosmos Predict 2.5, a robot video generation model, using LoRA/DoRA. The method reduces GPU resource requirements while maintaining generation quality for specialized robotics use cases.
PaddleOCR 3.5 integrates a Transformers backend for OCR and document parsing tasks. The new version improves accuracy and flexibility by leveraging Transformers models, enabling better text recognition and structured data extraction.
Hugging Face launches a public leaderboard to evaluate open-source AI agents. The platform ranks models by their ability to complete complex tasks, with reproducible benchmarks and transparent results.
IBM and Hugging Face release Granite Embedding Multilingual R2, an open-source embedding model under Apache 2.0 license. The model supports 32K token context and delivers best-in-class retrieval quality for sub-100M parameter models across multiple languages.
Hugging Face introduces an asynchronicity technique for optimizing continuous batching in inference servers. The method improves throughput by handling requests non-blockingly, reducing latency and increasing GPU resource utilization.
Hugging Face and AWS collaborate to provide optimized building blocks for foundation model training and inference on AWS infrastructure, including SageMaker integrations and open-source tools.
Hugging Face introduces EMO, a pretrained mixture of experts (MoE) model designed to develop emergent modularity. The approach aims to create specialized experts that naturally form during training, improving model efficiency and performance.
vLLM transitions from v0 to v1 prioritizing correctness before optimizations. The update introduces reliability and accuracy improvements in LLM inference, focusing on result validation before applying reinforcement learning techniques.
Hugging Face adds anti-Benchmaxxer filtering to the open ASR leaderboard to prevent artificial benchmark optimization. The system detects models over-optimized for test metrics without real generalization.