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 introduces multiple inference optimization techniques for Stable Diffusion XL (SDXL), including fp16 quantization and PyTorch 2.0's SDPA attention.
The Hugging Face team conducted a complete reproduction and technical deep dive into openai/lm-human-preferences, the RLHF codebase open-sourced by OpenAI in 2019.
Hugging Face has launched the Text Embedding Inference (TEI) service on Inference Endpoints, supporting efficient deployment of open-source text embedding models.
DALL·E 3 is now live in ChatGPT Plus and Enterprise. OpenAI stated that it has built a safety mitigation stack to support broader release and simultaneously updated progress on provenance research.
Hugging Face has released Gradio-Lite (@gradio/lite), a JavaScript library based on Pyodide that enables Gradio applications to run directly in the browser without a server.
PyTorch 2.1 has been officially released. Key new features include: torch.compile supports automatic dynamic shapes to reduce repeated compilations caused by tensor size changes; torch.distributed.checkpoint supports parallel saving/loading across multiple ranks and resharding under topology changes.
Hugging Face officially announced that the Diffusers library now supports JAX + Cloud TPU v5e, enabling high-performance, low-cost inference for Stable Diffusion XL (SDXL).
Hugging Face has introduced the chat_template attribute in tokenizers, using Jinja templates to store the dialogue format used during model training, thereby avoiding silent performance degradation caused by input format mismatches; it supports direct application via tokenizer.apply_chat_template() and provides three typical template examples, including ChatML.
Hugging Face released a tutorial guiding users on how to fork and configure the AI Comic Factory Space by setting LLM_ENGINE and RENDERING_ENGINE to INFERENCE_API.
A Hugging Face blog post introduces how to perform reinforcement learning fine-tuning on Stable Diffusion models using the Denoising Diffusion Policy Optimization (DDPO) method via the newly integrated DDPOTrainer.
Hugging Face released a tutorial for non-engineers demonstrating how to build a chat application based on LLaMA 2 7b in three steps using Spaces, AutoTrain, and ChatUI without code.
Mistral AI officially announced the release of its first open-source large language model, Mistral 7B. The model surpasses all currently available open-source models up to 13B parameters on standard English and code benchmarks. It is licensed under Apache 2.0, supporting unrestricted use. Developed using a proprietary MLops stack and data processing pipeline, the model is suitable for tasks such as summarization, structuring, and question answering, offering fast inference speed and low cost.
Mistral AI released the Mistral 7B (7.3B parameters) large language model, utilizing Grouped-query attention and Sliding Window Attention.
Hugging Face tested 60 configurations (including different instance types, concurrent request counts, and GPTQ quantization) for deploying Llama 2 (7B/13B/70B) on Amazon SageMaker. The tests measured latency (ms/token) and throughput (tokens/sec), providing specific recommendations for three scenarios: cost-effectiveness, highest throughput, and lowest latency. All code, raw data, and tables are open source.
ChatGPT has added multimodal capabilities, supporting visual input (seeing), voice input (listening), and voice output (speaking).
Hugging Face announced the launch of the "Inference for PROs" service for PRO users, providing dedicated HTTP API endpoints and higher rate limits for selected models such as Llama 3, Mixtral, Code Llama, and Stable Diffusion XL.
Hugging Face's official blog post "Optimizing your LLM in production" systematically explains three key technologies for improving the inference efficiency and memory efficiency of large language models.
Hugging Face releases the Würstchen diffusion image generation model, which employs two-stage compression (Stage A+B as Decoder, Stage C as Prior).
The Hugging Face blog introduces a complete practice for fine-tuning Llama 2 70B using PyTorch FSDP on two nodes with a total of 16 A100 80GB GPUs.
The Hugging Face blog systematically reviews the two quantization schemes natively supported in 🤗 Transformers: bitsandbytes and auto-gptq, covering their applicable scenarios.
Hugging Face has released SafeCoder, an enterprise-grade code assistant based on the open-source StarCoder model (15.5B parameters, supporting 80+ programming languages, with an 8192-token context window).
The Hugging Face blog announced the launch of the T2I-Adapter series of control models for Stable Diffusion XL (SDXL), supporting sketch and canny.
TII released Falcon 180B, with 180 billion parameters, pre-trained for a single epoch on 3.5 trillion tokens from the RefinedWeb dataset, making it the largest open-source language model at the time.
A Hugging Face blog post introduces how to reduce the inference time for AudioLDM 2 text-to-audio generation (10.24 seconds) from approximately 14 seconds to under 1 second while maintaining audio quality, using four optimization techniques: Flash Attention, float16 half-precision, torch.compile, and DPMSolverMultistepScheduler. It also demonstrates memory optimization tricks such as CPU offload to support long audio generation.