Hugging Face just acquired French startup Pollen Robotics to accelerate its push into embodied AI. The company plans to sell Pollen's humanoid robot, Reachy 2, while letting developers download and improve its open codebase. It marks a major step in bringing open-source collaboration directly to physical robot hardware.
Hugging Face just launched SmolVLM2, a family of lightweight multimodal models coming in 2.2B, 500M, and 256M parameters. The 500M version delivers video understanding comparable to the 2.2B model at a fraction of the memory footprint. With day-zero MLX and Swift support, these models can run completely offline on an iPhone or inside a free Google Colab.
Hugging Face just published a breakdown of AI agent autonomy levels and their ethical risks. The team warns that ceding control sharply increases safety hazards, specifically recommending developers avoid building fully autonomous agents that can write and execute their own code.
Hugging Face just launched the Synthetic Data Generator, a no-code tool that turns natural language descriptions into custom datasets. Powered by distilabel and free inference APIs, it crafts text classification and chat datasets in minutes. The output pushes directly to Argilla and the Hub, making dataset curation and fine-tuning painless for non-coders.
Hugging Face just made 83 popular open models available directly within Amazon Bedrock Marketplace. The models are managed via SageMaker JumpStart under the hood, allowing developers to plug open weights into Bedrock features like Agents, Knowledge Bases, and Guardrails with standard Bedrock API calls.
A group called Sora PR Puppets just leaked access to OpenAI's Sora on Hugging Face, briefly letting the public generate 1080p video clips before getting shut down. The collective claims OpenAI exploited early testers for unpaid labor and PR hype rather than genuine artistic collaboration. Why it matters: Creator backlash and delays are making OpenAI's video ambitions significantly harder.
NVIDIA just released a major suite of robotics development tools at CoRL, headlined by the open-source launch of Isaac Lab 1.2 and six new Project GR00T workflows for humanoids. The company also open-sourced its Cosmos visual tokenizer, running up to 12x faster to accelerate world model training. Why it matters: NVIDIA is actively trying to cement its compute and software stack as the unavoidable operating backbone for the next generation of physical AI.
Recraft just launched Recraft V3, claiming the #1 spot on Artificial Analysis's text-to-image leaderboard with an 1172 ELO rating, surpassing Midjourney and OpenAI. The new model lets creators place and scale exact text within designs, handles long-form copy, and natively generates vector graphics. It is available immediately for free and paid tiers across web, mobile, and API.
Hugging Face just acquired XetHub, a Seattle startup founded by ex-Apple engineers to overhaul how massive AI models and datasets are stored. Instead of relying on Git LFS and re-uploading gigabytes for minor edits, XetHub’s chunking tech lets users sync just a few kilobytes. Why it matters: as models head toward trillions of parameters, version control needs to keep up.
Stability AI just dropped the research paper behind Stable Audio Open, detailing the architecture of its open-weights text-to-audio model. Trained on nearly 500,000 Creative Commons recordings, the model synthesizes up to 47 seconds of 44.1kHz stereo sound and foley effects directly on consumer GPUs. The weights are live on Hugging Face for non-commercial use and eligible small businesses.
Apple's research team just released open DCLM models with 7B and 1.4B parameters on Hugging Face, sharing weights, code, and training datasets. The 7B model scored 63.7% on MMLU, outperforming Mistral-7B while using 40% less compute. It shows Apple's data curation prowess, though it's research-focused rather than on-device software.
Hugging Face just brought Google Cloud TPU v5e chips to its Inference Endpoints and Spaces, starting at $1.38 an hour. Powered by the open-source Optimum TPU library, developers can deploy and serve open models like Gemma, Llama, and Mistral in a few clicks. It gives builders a viable, budget-friendly alternative to scarce Nvidia GPUs.
Shanghai AI Lab just open-sourced InternLM 2.5-7B, cramming a massive 1M token context window into a lightweight architecture. The model claims top scores for sub-12B models on Hugging Face's Open LLM Leaderboard, bringing heavy-duty reasoning and tool use to local hardware.
Apple just released 20 new Core ML models and four datasets on Hugging Face, including FastViT and DepthAnything. Built specifically to run on-device without an internet connection, they let developers build fast computer-vision features while keeping user data strictly private.
Hugging Face just launched LeRobot, an open-source robotics library designed to act as the "Transformers for robotics." The toolkit includes pretrained models, simulation integrations, and lightweight dataset tools to control hardware from robotic arms to humanoids. Why it matters: Hugging Face is betting on open-source and crowdsourced data to prevent big tech from monopolizing physical AI.
Hugging Face just launched the Open CoT Leaderboard to measure how much chain-of-thought prompting actually boosts LLM accuracy, rather than tracking absolute benchmark scores. Early results from 30 models reveal small 7B models often gain more from step-by-step reasoning than giants, though CoT can occasionally backfire and degrade accuracy. It gives developers a real look at prompt effectiveness beyond raw capability.
Researchers just open-sourced JAT, a multi-purpose Transformer agent built to reproduce and improve upon Gato. Running on a single network, it tackles 157 tasks across Atari games, navigation, and robotic control, matching 65.8% of expert performance while supporting basic text and vision tasks. Why it matters: It opens up generalist embodied AI research with fully public models and datasets.
Meta just launched Llama 3, delivering 8B and 70B open-access models trained on over 15 trillion tokens. The smaller 8B variant can easily run and fine-tune on consumer GPUs down to 4-bit precision, backed by an expanded 128k vocabulary. Why it matters: Meta is resetting the benchmark for open AI, giving developers frontier-level capabilities without proprietary API lock-in.
Hugging Face just released Idefics2, an open-source 8B vision-language model licensed under Apache 2.0. It preserves native image resolutions, significantly upgrades OCR, and punches way above its weight class on visual QA benchmarks against 30B+ rivals. Why it matters: Developers get an efficient multimodal base model ready to fine-tune directly in Transformers.
Google just released CodeGemma in collaboration with Hugging Face, bringing a family of open-access code specialists built on Gemma. Trained on an extra 500 billion tokens, the lineup includes an infilling-focused 2B model that runs on consumer hardware alongside 7B base and instruct variants. Why it matters: It offers developers fast, private code completion without needing massive cloud infrastructure.
