Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 5, 2026, 9:00 AM
The Hugging Face blog published a tutorial guiding users on how to export ComfyUI workflows (such as Nathan Shipley's Flux[dev] Redux + Depth workflow) into Python scripts using the ComfyUI-to-Python-Extension.
Vectara has released the Hughes Hallucination Evaluation Model (HHEM) evaluation leaderboard built using the Hugging Face leaderboard template. It is used to assess the hallucination rates of models such as GPT-4, Gemini, and Llama 2 in document summarization tasks. The leaderboard supports user submissions for new model requests and automatic result updates; the code is open-source and deployed as a Hugging Face Space.
OpenAI launches the GPT Store.
OpenAI has launched the ChatGPT Team subscription plan for teams of all sizes, providing a secure and collaborative workspace to enhance the effectiveness of ChatGPT in work scenarios.
A Hugging Face blog post introduces the community-developed Unsloth library, which rewrites PyTorch modules using Triton kernels to boost LLM fine-tuning speeds by up to 2.7x and reduce memory usage by up to 74% without compromising QLoRA accuracy; it supports architectures such as Llama and Mistral, along with TRL's SFTTrainer and DPOTrainer, and offers pre-quantized 4-bit models and multiple free Colab runnable examples.
Hugging Face has released aMUSEd, an efficient non-diffusion text-to-image model based on Masked Image Modeling (MIM), which is an open-source reproduction of Google's MUSE. The model has approximately 800M parameters, supports text-to-image generation and zero-shot image inpainting, and is fully integrated into diffusers. It employs the CLIP-L/14 text encoder and U-ViT architecture, supports micro-conditioning control, and uses the OpenRAIL license.
Hugging Face has released an advanced Dreambooth LoRA training script for Stable Diffusion XL, integrating pivotal tuning (including textual inversion).
A Hugging Face blog post introduces how to use Speculative Decoding to improve the inference speed of large Whisper models for speech transcription by approximately 2x while ensuring output consistency.
Mistral AI released Mixtral 8x7B, an open-weight sparse mixture-of-experts model (SMoE) with a total of 46.7B parameters, using 12.9B parameters per token, supporting a 32k context window, covering five languages: English, French, Italian, German, and Spanish. It surpasses Llama 2 70B on most benchmarks and is 6 times faster in inference. The instruction-tuned version, Mixtral 8x7B Instruct, scores 8.30 on MT-Bench.
Mistral AI launched the platform service La Plateforme, opening up mistral-tiny (Mistral 7B Instruct v0.2, English only, MT-Bench 7.6).
Mistral releases the Mixtral 8x7B open-source large language model, which adopts a sparse Mixture of Experts architecture with approximately 45B parameters and a 32k context window, supporting five languages including English and possessing coding capabilities.
Hugging Face has released the Optimum-NVIDIA inference library, which enables up to 28x throughput improvement and 3.3x reduction in time-to-first-token latency on NVIDIA platforms via a single-line code replacement; the library initially supports float8 formats for Ada Lovelace and Hopper architectures, is compatible with the LLaMA model series (including fine-tuned versions), and offers two usage methods: pipeline and Model API.
Hugging Face and AMD announced a partnership to provide Hugging Face Transformers support for AMD Instinct MI210/MI250/MI300 series GPUs.
Hugging Face has implemented dynamic loading and unloading of Stable Diffusion LoRA adapters in its Inference API, keeping the base model resident.
After adding the DROP evaluation to the Open LLM Leaderboard, Hugging Face found that most models scored below 10 on F1. In-depth investigation revealed that the normalization step failed to handle numbers following line breaks, and the use of English periods as generation terminators led to truncation of floating-point numbers and long answers. After confirming the root cause with the Zeno and EleutherAI teams, they decided to temporarily remove DROP from the leaderboard pending a corrected evaluation implementation.
Sam Altman returns as Chief Executive Officer of OpenAI, Mira Murati serves as Chief Technology Officer, Greg Brockman returns as President, and OpenAI establishes a new initial board.
Hugging Face has integrated Latent Consistency LoRA (LCM LoRA) into the diffusers library, enabling models such as SDXL and Stable Diffusion v1.5 to generate high-quality images in just 4 steps.
Hugging Face's official blog announced the deployment of Llama 2 text generation on AWS Inferentia2 accelerators via optimum-neuron. It offers pre-compiled models (7B/13B, including budget/latency/throughput configurations), complete code examples for export and inference, and publishes measured data on compilation time, end-to-end latency, and throughput.
Explosion has launched the Prodigy-HF plugin, enabling direct integration between Prodigy and the Hugging Face ecosystem. It supports fine-tuning any transformer model from the Hugging Face Hub (such as distilbert-base-uncased) via the command line for NER and text classification tasks.
OpenAI has released GPTs, enabling users to create customized versions of ChatGPT by combining instructions, additional knowledge, and any skill sets.
At DevDay, OpenAI released GPT-4 Turbo (supporting 128K context and price reductions), GPT-4 Turbo with Vision, the DALL·E 3 API, and the new Assistants API.
Hugging Face Hub has introduced a storage region feature for the Enterprise Hub plan, supporting the specification of US or EU regions for storing an organization's models and datasets; the Asia-Pacific region will be launched subsequently. This feature meets GDPR compliance requirements and provides European users with approximately 4–5 times faster upload and download speeds when storing in the EU region.
The official Hugging Face blog released a tutorial introducing how to use StarCoder series models (such as bigcode/starcoder, starcoderbase-1b) to fine-tune a dedicated coding assistant named HugCoder on Hugging Face's own GitHub codebases.
The Renumics Spotlight tool supports interactive visual exploration of Hugging Face datasets via the single line of code `spotlight.show(ds)`, automatically identifying data semantics; it supports loading model outputs such as prediction results and embedding vectors to locate failure samples and key clusters, and offers GUI layout customization and Python API integration capabilities.