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 official Hugging Face blog published a tutorial demonstrating how to successfully run DeepFloyd's open-source pixel-level text-to-image model IF-I-XL-v1.0 on free-tier Google Colab (T4 GPU + 13GB CPU RAM) using 🧨 diffusers.
ChatGPT users can now disable chat history and choose which conversations are available for model training.
Hugging Face's official blog published a tutorial introducing how to build Unity games as WebGL and deploy them to Hugging Face Spaces. It requires selecting the Static HTML template, configuring the WebGL build target, disabling compression formats, enabling Git LFS tracking for large files, and finally pushing the build files.
Hugging Face's official blog announced a collaboration with AWS to deeply adapt the Hugging Face Transformers framework to AWS Inferentia2 inference acceleration chips.
Hugging Face released a StackLLaMA tutorial that demonstrates how to perform Supervised Fine-Tuning (SFT), Reward Modeling (RM), and Reinforcement Learning optimization (PPO) on LLaMA-7B using RLHF. It builds a question-answering model using the StackExchange dataset; all code, PEFT configurations, 8-bit loading solutions, and training techniques are open-sourced in the Hugging Face TRL library and 🤗 Hub.
A Hugging Face blog post introduces how to use the 🤗 Optimum Habana library to deploy BLOOMZ (176B) and BLOOMZ-7B models for inference on Habana Gaudi2 accelerators.
The Hugging Face blog introduces how to accelerate Stable Diffusion inference on Intel Sapphire Rapids CPUs (such as Amazon EC2 r7iz.metal-16xl).
This tutorial demonstrates how to combine the Hugging Face and Flower frameworks to perform federated fine-tuning of the distilBERT model across multiple clients for the IMDB movie review sentiment classification task; it includes full code implementation, supports local simulation of the federated training process via Flower's simulation feature in Colab, and provides methods for building PyTorch-based clients and servers.
The Hugging Face blog details a three-step process for training ControlNet using the diffusers library: planning control conditions (such as facial keypoints) and building datasets (based on synthetic face data from FaceSynthetics).
OpenAI has implemented initial support for plugins in ChatGPT. Plugins are tools designed specifically for language models, with security as a core principle, helping ChatGPT access up-to-date information, perform calculations, or call third-party services.
Hugging Face Hub has added support for rendering Jupyter notebooks (.ipynb), displaying them directly in a human-readable format and enabling one-click opening in Google Colab. This initiative aims to improve the reproducibility and educational value of models and datasets; currently, the Hub hosts over 7,000 notebooks.
PyTorch officially released version 2.0, with core updates including the stable torch.compile API, Beta distributed tensor DTensor, and Metal Performance Shaders (MPS) backend support for macOS GPU acceleration.
OpenAI released GPT-4, a new milestone in scaling deep learning. It is a large multimodal model supporting image and text inputs with text output, achieving human-level performance on multiple professional and academic benchmark tests, but still underperforming humans in many real-world scenarios.
Hugging Face officially released the integration between the TRL and PEFT libraries, making Reinforcement Learning from Human Feedback (RLHF) fine-tuning of Large Language Models (LLMs) more accessible; this solution loads models using 8-bit matrix multiplication.
Kakao Brain and Hugging Face jointly released ViT and ALIGN models trained on the open-source COYO-700M image-text dataset, marking the first publicly available implementation of ALIGN.
The official Hugging Face blog announced the formal integration of ControlNet into the diffusers library, releasing the StableDiffusionControlNetPipeline.
Following the February 2023 Turkey earthquake, a volunteer team rapidly built the afetharita disaster map application based on Hugging Face Hub, using easyocr + Gradio to implement screenshot text extraction.
Hugging Face has released version 1.1 of Swift 🧨Diffusers, a native Mac app accelerated by Core ML that supports models such as Stable Diffusion 1.5/2/2.1.
Hugging Face has introduced the Inference Endpoints managed service, supporting one-click deployment of models from the Hub (including transformer and sklearn models) to various instances (including GPU/CPU) on cloud platforms such as AWS and Azure.
Salesforce Research releases BLIP-2, a zero-shot vision-language model that bridges frozen visual encoders (ViT) and frozen large language models (OPT/Flan T5) via a lightweight Querying Transformer (Q-Former); it supports image captioning, prompt-based image captioning, visual question answering, and ChatGPT-like multi-turn dialogue, and has been open-sourced in 🤗 Transformers with pre-trained checkpoints provided.
Hugging Face releases the 🤗 PEFT library, offering parameter-efficient fine-tuning methods such as LoRA, Prefix Tuning, Prompt Tuning, and P-Tuning.
Hugging Face officially announced that SpeechT5 has been formally integrated into the 🤗 Transformers open-source library. The model was proposed by Microsoft Research Asia and supports three major tasks: Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Speech-to-Speech (voice conversion).
Hugging Face has released the open-source tool AI vs. AI, designed to evaluate the relative strength of reinforcement learning models in multi-agent environments through model battles and an ELO rating system. The system consists of three parts: a Spaces matchmaking service, a Hub Dataset for storing results, and a Leaderboard for displaying rankings. It has currently launched its first challenge in the SoccerTwos environment, with 48 models participating.
OpenAI has launched the ChatGPT Plus subscription service, targeting users who seek more stable access, faster responses, and priority experience with new features. This service is based on ChatGPT, an AI that supports conversation, follow-up questions, and challenging incorrect assumptions.