AI products · OpenAI News
OpenAI Releases ChatGPT Enterprise ↗
OpenAI released ChatGPT Enterprise, offering enterprise-grade security and privacy protections, as well as the most powerful version of ChatGPT available at the time.
TIAOTECH / NEWS
Stay curious. Stay in the loop.
Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 5, 2026, 9:00 AM
AI models · Hugging Face Blog
Meta 发布 Code Llama：基于 Llama 2 的开源代码大语言模型家族 ↗
Meta 发布 Code Llama，一个基于 Llama 2 的开源代码大语言模型家族，包含 7B、13B 和 34B 三种参数规模，以及 base、Python-specialized 和 instruction-tuned 三类变体。
AI products · Hugging Face Blog
Hugging Face Integrates AutoGPTQ into the Transformers Library ↗
Hugging Face has officially integrated AutoGPTQ into the Transformers library, supporting 8/4/3/2-bit quantization of large language models using the GPTQ algorithm.
AI products · OpenAI News
OpenAI Opens GPT-3.5 Turbo Fine-Tuning ↗
Developers can now use their own data to fine-tune GPT-3.5 Turbo to suit their specific use cases.
AI products · Hugging Face Blog
Hugging Face Releases SafeCoder: A Self-Hosted Commercial Code Assistant Solution for Enterprises ↗
Hugging Face announced the launch of SafeCoder, a self-hosted commercial code assistant solution for enterprises. Built on the open-source StarCoder model family, it supports deployment on customer-owned infrastructure (including various hardware such as NVIDIA, AMD, Intel, AWS, and Habana). Training and inference occur entirely within the VPC, ensuring code never leaves the environment. It provides containerized, hardware-accelerated inference services and supports IDE plugins including VSCode and IntelliJ.
AI models · Hugging Face Blog
Hugging Face Releases IDEFICS: Open-Source Multimodal Vision-Language Model ↗
Hugging Face has released IDEFICS, an open-source vision-language model built upon public data and models (LLaMA v1 and OpenCLIP), offering two parameter sizes (9B and 80B) along with corresponding instruction-tuned versions.
Practical tips · Hugging Face Blog
Hugging Face Blog Tutorial: Deploying the DeepFloyd IF Text-to-Image Model with BentoML ↗
The official Hugging Face blog published a tutorial demonstrating how to use the open-source framework BentoML to deploy DeepFloyd IF (an open-source, pixel-space, three-stage cascaded text-to-image model) as a production service.
Practical tips · Hugging Face Blog
Hugging Face Tutorial: Optimizing the Bark Text-to-Speech Model with 🤗 Transformers ↗
The official Hugging Face blog published a tutorial demonstrating how to perform three optimizations on the Bark text-to-speech model using the 🤗 Transformers, Optimum, and Accelerate libraries.
Practical tips · Hugging Face Blog
Hugging Face Releases Swift Transformers: Enables Local Execution of Large Language Models on Apple Devices ↗
Hugging Face has released the Swift Transformers toolchain, which includes swift-transformers (a Swift implementation of transformers-like APIs), swift-chat (a demo application), exporters, and transformers-to-coreml (Core ML conversion tools). It also provides pre-converted models such as Llama 2 7B and Falcon 7B.
Practical tips · Hugging Face Blog
Hugging Face Blog Details DPO Method and Releases TRL Library Support ↗
The Hugging Face blog introduces the Direct Preference Optimization (DPO) method, which transforms the reinforcement learning objective in RLHF into a binary cross-entropy loss that can be directly optimized.
Practical tips · Hugging Face Blog
Hugging Face Releases Tutorial on Deploying MusicGen with Inference Endpoints ↗
The Hugging Face blog introduces how to deploy the MusicGen model using Inference Endpoints, including copying the facebook/musicgen-large repository, writing a custom inference function in handler.py and requirements.txt, creating endpoints, and providing curl and Python call examples along with audio generation code.
Practical tips · Hugging Face Blog
Hugging Face and Zama Collaborate to Explore Fully Homomorphic Encryption for Large Language Models ↗
In a blog post, Hugging Face and Zama showcased how to use Fully Homomorphic Encryption (FHE) to protect user data privacy and model intellectual property during LLM inference, with the specific implementation based on modifying GPT2 using the transformers library.
Practical tips · Hugging Face Blog
Huggy Lingo: Automatically Completing Language Metadata for Hugging Face Hub Datasets Using Machine Learning ↗
Hugging Face introduced the Huggy Lingo method, which uses Meta's facebook/fasttext-language-identification model to predict languages for datasets lacking language labels.
Practical tips · Hugging Face Blog
Hugging Face Blog Publishes Tutorial on PS1-Style 3D Asset Generation ↗
The Hugging Face blog published "Practical 3D Asset Generation: A Step-by-Step Guide," detailing how to use Shap-E (an open-source diffusion model by OpenAI) to generate text-to-3D models.
AI models · Hugging Face Blog
Segmind Open-Sources SD-Small and SD-Tiny Distilled Models ↗
Segmind has open-sourced Stable Diffusion variants SD-Small and SD-Tiny, compressed via knowledge distillation (KD), which reduce the number of parameters by 35% and 55%, respectively, compared to the baseline model.
Practical tips · Hugging Face Blog
Hugging Face and Apple Release Core ML Version of Stable Diffusion XL with Mixed-Bit Palette Quantization Scheme ↗
Hugging Face and Apple have collaborated to release a Core ML implementation of Stable Diffusion XL.
AI products · Hugging Face Blog
Hugging Face Releases Agents.js: An LLM Tool-Calling Library for JavaScript Environments ↗
Hugging Face has released the Agents.js library, which supports integrating tools with LLMs in browsers or Node.js. It includes built-in multimodal tools and allows for extensible custom LLMs and tools; it offers the HfAgent API.
AI products · OpenAI News
ChatGPT Launches Custom Instructions Feature ↗
OpenAI has introduced a custom instructions feature for ChatGPT, enabling users to set preferences that ChatGPT will consistently follow in all subsequent conversations.
Practical tips · Hugging Face Blog
🤗 Diffusers Library One-Year Anniversary: New Features Including Video, 3D, LoRA, SDXL, and More ↗
Hugging Face announced that the 🤗 Diffusers open-source library is celebrating its one-year anniversary, with new support for higher-quality image generation models such as DeepFloyd IF and Stable Diffusion XL (SDXL).
Practical tips · Hugging Face Blog
Hugging Face Fully Supports the Release of Meta's Llama 2 Open-Source Large Language Model ↗
Meta released the Llama 2 model family (7B/13B/70B parameters), including pre-trained versions and RLHF-fine-tuned Llama 2-Chat versions, under a commercially usable license; Hugging Face simultaneously launched all 12 open-source models on its official platform.
Practical tips · Hugging Face Blog
Overview of Hugging Face's Open-Source Text Generation and Large Language Model Ecosystem ↗
The Hugging Face blog systematically introduces open-source text generation models, including causal language models such as Llama 2, Falcon, and MPT-30B, and text-to-text models such as FLAN-T5.
Practical tips · Hugging Face Blog
Hugging Face Releases Experimental AI WebTV Video Streaming Project ↗
Hugging Face has launched the experimental AI WebTV project, which generates videos using Zeroscope V2 (zeroscope_v2_576 and zeroscope_v2_XL).
Practical tips · Hugging Face Blog
Hugging Face Releases Complete Tutorial for Fine-Tuning Stable Diffusion on Intel CPUs ↗
A Hugging Face blog post released a tutorial demonstrating how to fine-tune the Stable Diffusion model on clusters equipped with Intel Sapphire Rapids CPUs, using textual inversion techniques requiring only 5 example images; the tutorial covers the full workflow including server configuration, installation of oneCCL and IPEX, diffusers code modifications, distributed training commands, and OpenVINO-accelerated inference.
Practical tips · Hugging Face Blog
Building a Real-Time ML Drawing Game, Doodle Dash, in the Browser with Transformers.js ↗
The author fine-tuned the apple/mobilevit-small model based on the Google Quick, Draw! dataset, converted it to ONNX format, and ran it in real-time in the browser via Transformers.js, implementing the drawing recognition game Doodle Dash with over 60 predictions per second; the tutorial covers three parts: training, deployment, and game design, and provides links to the Colab notebook, Hugging Face model page, and GitHub source code.