Hugging Face

NVIDIA Cosmos 3 is the first open-source omni-model for physical AI, integrating world generation, physical reasoning, and action generation into one unified system. Built on MoT architecture, it supports robotics, autonomous driving, and synthetic data pipelines via Hugging Face and Diffusers.

JetBrains releases Mellum2, a 12B-parameter MoE model activating only 2.5B params per token, offering 2x+ faster inference than peers, optimized for text/code tasks and private/RAG deployments.

Gemma-4 12B is an encoder-free, unified multimodal model that runs directly on laptops with 16GB VRAM. It matches the performance of the 26B MOE with less than half the memory footprint, ships with Hermes and agent tools, macOS Edge Gallery, and RTLM, and is released under Apache 2.0.

Gemma 4 12B is a unified, encoder-free multimodal model bringing high-performance multimodal intelligence to your laptop. It matches the performance of our 26B MoE at less than half the memory footprint, supports native audio inputs, and runs locally on 16GB VRAM hardware with low-latency multi-step reasoning.

Hugging Face introduces MCP tool integration, enabling the Reachy Mini robot to use remote tools from Hugging Face Spaces via MCP, such as weather lookup and web search, without modifying the app. Tools run in the cloud, users can publish their own tools, and a single command adds new capabilities like asking for Paris weather today.

Holo3.1 is Hugging Face's new computer-use agent model supporting cross-platform, multi-framework deployment and first releasing quantized weights (FP8/Q4 GGUF/NVFP4) for local inference.

Mellum2 is an open-source 12B parameter AI model from JetBrains, using MoE architecture to activate only 2.5B parameters per token, reducing inference time by over 50% compared to similar-sized models, specifically designed for software engineering environments with applications in routing, RAG pipelines, and private AI deployment.

LeLab is a GUI tool for robot control, enabling code-free dataset creation and model training.

NVIDIA's synthetic data generation agent compresses visual inspection data collection from months to hours, generating PCB, glass, and metal defect images via natural language prompts to boost model training efficiency.

This paper proposes a new method using language models to predict GPU kernel runtime performance, demonstrating superior results compared to traditional methods in multiple benchmarks.

The open-weight model MiniMax M3 has reached performance comparable to GPT-5.5 and Opus 4.7, outperforming Gemini 3.1 Pro in coding tasks, and costs 10x less to use, with weights to be released on Hugging Face next week.

Tensors are the core data structure in machine learning that convert real-world inputs like text, images, and audio into numerical matrices for computation, producing meaningful outputs. They are simply numbers organized by shape, including scalars (0D), vectors (1D), matrices (2D), and higher-dimensional arrays.

Ideogram released open weights for its v4 image model, enabling local deployment and fine-tuning as the state-of-the-art open-source text-to-image model.

Tensors are fundamental structures in machine learning that organize numbers, enabling models to process real-world data like text, images, and audio by converting them into numerical forms for the complete input-to-output data flow.

Transformers.js implements efficient machine learning inference in JavaScript using ONNX, automatically handling model download, caching, and tensor conversion to simplify Web AI integration.

Google announces its model weights are compatible with major open-source ecosystems and can be directly downloaded from Hugging Face and Kaggle, lowering deployment barriers.

NVIDIA announces that Cosmos 3 is fully open-sourced, including model weights and post-training recipes, now available on Hugging Face for developers to use and research freely.

NVIDIA has fully open-sourced Nemotron 3 Ultra, including weights, synthetic data, and post-training recipes, now available on Hugging Face.

Ideogram released v4 open-weight image model supporting local fine-tuning and private deployment, though the source is a brief announcement lacking technical depth.

Hugging Face CEO recommends AI agents store traces privately on HF by default to enable history analysis, sharing, and post-training optimization.

NVIDIA's Physical AI Agent Skills tool enables video augmentation via simple prompts, but the article only demonstrates operational steps without technical depth or principle analysis, resulting in low information density.

Cohere sponsors a hackathon in collaboration with Hugging Face and Gradio, focusing on developing smaller, more cost-effective, and practical AI models.

Cohere posted a tweet announcing that registration for Hugging Face and Gradio will close tomorrow, and recommending the Transcribe and Tiny Aya models for the hackathon.

Nvidia, the 'King of American Open-source AI', has over 1,000 repositories on Hugging Face (820 models, 249 datasets, 57 spaces), and released new models like Cosmos 3 and Alphamayo 2 Super, adopting the OpenMDW framework to advance open-source AI.

Nvidia has over 1,000 public repositories (820 models, 249 datasets, 57 spaces) and nearly 60,000 followers on Hugging Face, with LocateAnything as the #1 trending model.

The Hugging Face blog has become a major source of learning and news about AI for the community, with NVIDIA, JetBrains, and others publishing content there.

The tweet calls for more attention to Arcee, a U.S.-based open-source AI model company, but lacks depth, novelty, and practicality, serving only as a simple recommendation.

This tweet only shares a Hugging Face link to Google's Gemma-4-12B-it model without technical analysis, benchmarks, or engineering guidance, offering minimal informational value for practitioners.

This tweet only shares a Hugging Face dataset link (nanoclaw-traces) without technical context or engineering value, making it unsuitable as reading material.

This tweet is merely a brief announcement for the Nemotron 3 Ultra model release, lacking technical details, architecture specs, or performance data, offering no engineering reference value.