Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Hugging Face and KerasHub announced an integration that enables KerasHub users to directly load Transformers models from the Hugging Face Hub (initially supporting Gemma 1/2).
Hugging Face announced support for Google TPU v5e in its Inference Endpoints and Spaces services, offering three instance configurations: v5litepod-1/4/8.
Hugging Face Dataset Hub has launched four new dataset search features: filtering by modality (Text/Image/Audio/Tabular/Time-Series/3D/Video/Geospatial).
Hugging Face's official blog announced that the transformers.agents framework, originally used for building agents, has been upgraded to a standalone library called smolagents, with highly similar APIs between the two.
The deepseek-chat model has been upgraded to DeepSeek-V2-0628. HumanEval Pass@1 improved from 79.88% to 84.76%, MATH ACC@1 improved from 55.02% to 71.02%, BBH improved from 78.56% to 83.40%, and the win rate against GPT-4-0314 on Arena-Hard improved from 41.6% to 68.3%. Role-playing capabilities have been significantly enhanced.
Google released Gemma 2, including two base models, gemma-2-9b and gemma-2-27b, along with their instruction-tuned versions; the context window is 8K tokens. The 9B model was pre-trained on approximately 8 trillion tokens.
Hugging Face released a tutorial demonstrating how to fine-tune Florence-2, a multimodal model launched by Microsoft in June 2024, on the DocVQA dataset. The model's Levenshtein similarity on the DocVQA validation set was 0 before fine-tuning and improved to 57.0 after 7 epochs of fine-tuning. The tutorial includes complete Colab code, configurations adapted for A100/T4 GPUs, a vision encoder freezing strategy, and links to inference demonstrations.
OpenAI acquires Rockset.
The Hugging Face team released BigCodeBench, a new code generation benchmark containing 1,140 function-level tasks that cover diverse function calls across 139 libraries, with each task averaging 5.6 test cases.
The Hugging Face Accelerate blog details behavioral differences between PyTorch FSDP and DeepSpeed in mixed-precision training: DeepSpeed defaults to promoting optimizer parameters to float32, while FSDP natively preserves torch_dtype; this difference leads to varying loss convergence performance under the same learning rate.
Hugging Face has released the RLOO (REINFORCE Leave One-Out) trainer in the TRL library as a lightweight online RLHF alternative to PPO.
Stability AI has released Stable Diffusion 3 Medium (2B parameters), which is now available on Hugging Face Hub and supported by Diffusers.
OpenAI and Apple have announced a partnership to integrate ChatGPT into Apple product experiences.
Hugging Face announced the official launch of the Hugging Face Embedding Container for Amazon SageMaker, which is built on Text Embedding Inference (TEI).
Hugging Face and Artificial Analysis jointly released the Text to Image Leaderboard and Arena, calculating ELO scores based on over 45,000 human image preference data points. The evaluation covers mainstream open-source and closed-source text-to-image models such as Midjourney, DALL·E, Stable Diffusion, and Playground; users can participate in voting and receive personalized model rankings.
Mistral AI announced the launch of model customization features on the la Plateforme platform, offering three methods: an open-source fine-tuning SDK (mistral-finetune.
Hugging Face, in collaboration with Cubzh and Gigax, has released NPC-Playground, enabling interaction with LLM-driven 3D agent NPCs directly in the browser; it supports teaching NPCs new skills using just a few lines of Lua script.
Hugging Face officially supports Assisted Generation for Intel Gaudi processors within Optimum Habana.
Hugging Face officially disclosed a recent unauthorized access incident on its Spaces platform, where some Spaces secrets may have been accessed. The company has revoked relevant HF tokens, encouraged users to switch to fine-grained access tokens, removed org tokens from infrastructure, integrated KMS, enhanced secret leak detection and automatic invalidation capabilities, and reported the incident to law enforcement and data protection authorities.
Mistral has released its first code large model, Codestral, featuring 22B parameters, a 32k context window, and support for over 80 programming languages. It offers two usage options: open-source weights (under the Mistral AI Non-Production License) and commercial licensing.
Mistral AI has released the Mistral Non-Production License (MNPL), which allows developers to use the technology for non-commercial purposes and research; the Codestral model is being released under this new license effective immediately; the company also stated that it will continue to release some models and code under Apache 2.0, forming a dual-license approach.
Hugging Face's official blog introduced its benchmarking tool for Text Generation Inference (TGI), which is used to quantitatively analyze the impact of different configurations (especially batch size) on prefill latency (TTFT).
The official Hugging Face blog released a complete guide to training and fine-tuning embedding models with the Sentence Transformers library, covering dataset loading (supporting formats such as Hugging Face Hub and local CSV/JSON/Parquet).
The Technology Innovation Institute (TII) of the UAE released Falcon 2-11B, the first model in the Falcon 2 series. It includes an 11B-parameter pure-text large language model and a vision-language model (VLM) built upon it, supporting English and ten languages including Spanish and French. The training data exceeds 5000B tokens, with a context window of 8192, under an Apache 2.0-derived license.