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
Hugging Face has switched its Inference Providers router for language models to OpenRouter's, so requests are now routed by live price, speed and reliability and retried on another provider if one fails. Pricing and code stay the same, but custom provider API keys no longer work for language models.

A Hugging Face walkthrough shows how to build a beverage can pick-and-place simulation in Isaac Sim 6.2 using SimReady assets: the UR17G arm, Robotiq 2F-140 gripper and a can, assembled and driven by agent-generated Python code.

IBM's Granite 4.1 3B and 8B models now have experimental LLMON previews that use reserved special-token delimiters to separate instructions from data, cutting indirect prompt-injection success rates close to zero on several benchmarks while matching standard Granite 4.1 Instruct performance.

Hugging Face published a guide with fine-tuning scripts and Docker containers for adapting the TRELLIS.2 3D generation model to a custom object category using LoRA, runnable locally or on Hugging Face Jobs. A footwear experiment trained roughly 9.3 million adapter parameters on 179 objects.

Tether AI Research has released BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for on-device tool calling, which roughly doubles its base model's agentic benchmark scores and beats a sibling 2.5x its size on five of six tests. Weights are open under Apache 2.0 in bf16, Q8 and Q4, as safetensors and GGUF.

Adding a pooling-aware loss term to Sentence Transformers training raises NanoBEIR@13 mean nDCG@10 retention to 97.8% at a 4x reduction in document vectors, versus 94.6% for a normally trained model.

Hugging Face's Transformers library now includes DecisionPipeline, AutoModelForDecision, and DecisionConfig for a new class of "decision models" that return probability distributions over predefined answers instead of generating text. The models support three primitives: choice, score, and noul, and can be served via a local /v1/systemone endpoint.

Elith and NVIDIA released Nemotron-Personas-Finance-Japan, a CC BY 4.0 synthetic persona dataset that pairs Japan's official statistics with household finance and financial literacy data, at 9M personas and roughly 1.5B tokens.

NVIDIA added four open-weight Cosmos 3 Nano checkpoints (16B trainable parameters each) under OpenMDW1.1 for driving-scene generation, streaming simulation, simulator-to-video rendering and bimanual robot video prediction.

Gradium's Phonon, a 31M-parameter on-device TTS model released two weeks ago under Apache 2.0 for English, tops the Open TTS Leaderboard's CPU streaming time-to-first-audio benchmark and ranks first in German and Spanish intelligibility.

IBM released Granite Switch 4.1.1, a backward-compatible patch that updates the Granite Speech ASR model to cover 10 languages with native punctuation and capitalization, plus an improved Decision adapter, without retraining or touching the base model.

Mistral Large 4, nicknamed Le Chonk, is a multimodal Mixture-of-Experts model with roughly 1T total and 49B active parameters and a 1M-token context window, released in FP8 and NVFP4 variants under a modified MIT-style open-weight license.

Hugging Face and partners launched Project ProtoStar, an initiative to help publicly funded research get shared as AI-ready datasets, models and code, with a first open challenge hosted by Carnegie Mellon University running November 2–13, 2026.

Two new Swift packages, AnySpeechRecognizer 0.1.0 and AnyDecisionModel 0.6.0, bring pluggable speech recognition and probability-based decision models to Apple platforms, with open models downloadable from the Hugging Face Hub and run locally.

NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that handles transcription in 40 languages, speaker diarization in 13, translation for 31, and audio question answering from a single checkpoint, alongside the Granary v2 data pipeline and fine-tuning recipes.

Liquid AI released LFM3, a family of six roughly 2B-parameter open-weight models built on a shared LFM3-1.7B-Base backbone, covering text, vision, audio, omni and decision use, with the text model outperforming the 35% larger LFM2.5-2.6B and fitting in about 1 GB at 4-bit precision.

Ai2's study finds LLM novelty judges are highly unstable: a small prompt wording change moved GPT-5.4's accuracy at picking the more novel of two ideas from 93.5% to 40.9%, and purpose-built evaluators underperformed a simple Claude Opus 4.6 prompt.

LitPivot, built by Hugging Face researchers with University of Pennsylvania collaborators, organizes retrieved literature around whichever part of a research proposal is being edited and offers critiques with supporting excerpts plus optional revisions. In a 17-person study, two blind domain experts rated ideas developed with it as better grounded in the literature than those made with a text editor and AI chat tool.

Microsoft and Hugging Face released ThinkingBox-Training, an open-source stack that reuses the same executable database-state checks used for grading agents as RL rewards. Fine-tuning Qwen3.8-27B with GRPO for 50 steps raised pass@1 from 51.70% to 60.78% and tasks solved on all 20 attempts from 38 to 107.

Hugging Face released OlmoEarth v1.3, a satellite embedding model whose 128-dimensional annual global embeddings score highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, alongside a free browser tool for exploring and classifying maps without a GPU.