AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
Hugging Face Blog
First-party releases and research from Hugging Face Blog. Headlines and excerpts link to the original articles.
Latest stories
Newest first
IBM has released Granite Timeseries Ensemble, an open recipe that combines four pre-trained Granite forecasting models into one call returning median and quantile forecasts. It ships with a uniform Probabilistic Ensemble as the default and an adaptive IQR alternative, requires no retraining or fitted weights, and both recipes beat all four members on aggregate GIFT-Eval MASE.

Hugging Face and Pollen Robotics published a tutorial for training a walking policy for the open-source bipedal robot Microduck in NVIDIA Isaac Lab 3.0 with the Newton physics backend, then exporting it as ONNX and running it on the physical robot. Policies can be shared on the Hugging Face Hub or submitted to the Microduck Arena Space.

A research team rebuilt the tokenizer of a production Portuguese essay grader built on Qwen3-4B using a method called VocabX, cutting measured inference cost by 77% without changing the base model, data or training recipe. The grader's average response fell from 350 to 101 tokens.

A new Hugging Face walkthrough shows how to use torch.profiler with the Neuron backend on AWS Trainium to find where inference time goes. In its RMSNorm example, fusing the operation with torch.rms_norm or torch.compile cut device time to roughly one-sixth of a naive multi-op version.

SimReady Central Library collects more than 2,000 simulation-ready OpenUSD assets on Hugging Face, free and without gating, spanning robot arms, grippers, warehouse equipment, racks and cooling infrastructure, with natural-language search powered by NVIDIA USD Search.

NVIDIA released Cosmos3-H-Dreams, a generative surgical simulator trained on clinical cholecystectomy recordings from the CMR Surgical Versius platform that runs interactively on a single RTX PRO 6000 GPU. Two checkpoints trade visual quality for inference cost via four-step and two-step denoising.

Allen Institute for AI (Ai2) released FairStress, a benchmark that tests whether LLMs resist persuasive pressure equally regardless of which demographic group a challenge favors, not just whether they pick the same candidate across identity swaps. It has 300 scenarios expanded to over 13 million test cases, plus a 50K-item FairStress-Core subset, and is built into Olmo-eval.

ALTK-Evolve, which appends learned behavioral guidelines to a stock agent's system prompt without changing model weights, reached 67.5% Pass^4 on the retail domain of the τ²-bench customer service benchmark using GPT-4.1, topping every retail entry in the October 12, 2026 leaderboard snapshot.

AnyLanguageModel 1.0 is out, a Swift package that keeps Apple's Foundation Models API but lets apps swap in local llama.cpp, MLX or Core ML models, or hosted providers like OpenAI, Anthropic, Gemini and Ollama, via a single import change. swift-huggingface 1.0 and swift-xet 1.0 ship alongside it.

Hugging Face announced TorchTPU, a native PyTorch backend that makes Google TPUs a regular device, so Transformers, Diffusers and Kernels now run on TPU by passing device_map="tpu" without torch_xla or Optimum TPU.

NVIDIA and AI4Bharat are releasing Nemotron-pretraining-Indic-v1, roughly 2.8 trillion tokens of Indic training text across 22 languages, alongside the open GPU-accelerated pipeline that generated it. The corpus splits into about 2,601.0 billion native-script tokens and 224.6 billion romanized tokens.

Reflection, a new US lab, released Beam on Hugging Face: an Apache 2.0 mixture-of-experts model with 501B total parameters and 23B active per token, a 1M-token context window, and day-0 support in transformers, vLLM, and SGLang.

IBM's Granite Switch 4.1 combines chat, decision-making and speech-to-text adapters into single 8B and 30B models that run on roughly the VRAM of the base model alone, all served from one hosted endpoint.

Apertus 1.5 adds native image and audio understanding, an optional thinking mode, a 262,144-token context window and improved tool use to the 8B and 70B models, with weights released July 24, 2026 under Apache 2.0.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model with roughly 816 billion parameters and about 42 billion active per token, supporting text, images, audio and video with a context window up to one million tokens.

HiRO-ACEv2 is an updated climate emulator that generates global 3 km weather data at roughly 1/300th the cost of the high-resolution X-SHiELD model while reproducing rare rain, wind and pressure extremes to within 3%.

Liquid AI released LFM2.5-24B-A2B, a 24B-parameter reasoning agent model with only 2B active parameters per token, designed to run agents locally on laptops and single GPUs. It scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B and decodes 184 tok/s on an Apple M5 Max.

AI2 replaced its priority-based GPU scheduler with a system built on GPU time budgets, hierarchical fair-share allocation and a time-slicing contract, reporting 98% of owed GPU hours delivered, 98% occupancy, and a 74% drop in repairs needing human intervention.

A Hugging Face blog post describes using the ML-intern agent in HuggingChat to build six custom models in a few days, including a 0.8B prompt rewriter distilled from a 9B teacher and trained for about USD 16 in compute. Each project ran from a single prompt, with the agent handling training, evaluation and Hub publishing.

Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M (experimental), built on its Liquid Foundation Models. d1-3B scores 48.57 on the Decision Index 0.2.1 and answers questions in under 50 ms on measured edge devices.