AI Briefing#hugging-face
The latest AI and developer news about #hugging-face, with the original source and a short summary.
The latest AI and developer news about #hugging-face, with the original source and a short summary.
NVIDIA, Pollen Robotics, and Hugging Face introduce a workflow to train walking policies for the Microduck robot in Isaac Lab 3.0 and deploy them via ONNX.
A Hugging Face blog post details system-level profiling methods for TorchNeuron on AWS Trainium, showing that operator fusion reduces RMSNorm device time from 2,355 µs to 375 µs.
The draft models combine a five-layer DFlash backbone with a learned bigram selector to propose multiple candidate continuations for target-model verification.
New 2B model ranks #2 on Open ASR Leaderboard while updated Turbo CTC achieves 19,870 RTFx.
The unified flow-matching model outperforms Proteina-Complexa by 5.8x in design yield while integrating sequence and all-atom structure generation.
The library provides pre-configured OpenUSD models with physics and thermal metadata to accelerate digital twin construction and robotics training.
The new model retains the Spark-X2.5-4B architecture while delivering state-of-the-art performance among similarly sized open-source models.
The model jointly optimizes task generation, scaffold construction, and solution rollouts via reinforcement learning to replace fixed human-curated training sets.
Transformers, Diffusers, and Kernels now run natively on TPUs using the TorchTPU backend, eliminating the need for torch_xla or JAX.
The 1.0 release adds support for OpenAI, Anthropic, and Gemini APIs alongside local models like Gemma 4 E2B and Qwen3.8 27B.
The new backend enables eager execution and native distributed APIs in Transformers, Diffusers, and Kernels, replacing the previous Optimum Neuron workflow.
The 4B parameter model uses a looped architecture with interleaved sliding-window and full attention to scale intelligence on a small footprint.
The new LoRA improves coordinate accuracy from 41% to 74% compared to the base model by returning page numbers and bounding boxes for every extracted field.
The model supports multimodal retrieval across 34 languages for text and 18 for images, with open commercial use rights.
The new 8B and 30B models integrate a Decision aLoRA, Guardian aLoRA, and the fastest ASR model on the OpenASR Leaderboard while maintaining base-model VRAM usage.
The 2B parameter model achieves up to 500 tokens per second locally in the browser on an M4 Max using custom WebGPU kernels.
The new V2 model achieves 92.33% accuracy on a sampled MMLU-Pro subset while optimizing reasoning token usage for agent workloads.
The new model supports text-to-image, multi-reference editing, and bounding-box layout control, with distilled checkpoints generating 1024x1024 images in ~2 seconds on H100 GPUs.
The open-weight model reduces thinking tokens to roughly half of the base model's usage while maintaining valid tool calls in agent loops.
The 3.66B parameter model scores 55.4 on a logic composite, matching Qwen3-8B at less than half the size.
The open-weight Aplomb 1 model achieves a KL divergence of 0.123 and a Brier score of 0.065 on the Typed Decisions leaderboard.
The 7B model generates 2K images and supports natural-language edits using an accelerated 8-step denoising schedule.
Achieves 0.743 accuracy by performing decisions with a single forward pass of a 27B model
Users can now chat with models hosted on Hugging Face Inference Providers by simply running `ssh chat.hf.co` without needing API keys or configuration.