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 has introduced a unified tool calling API across model families in the Transformers library, supporting models such as Mistral, Cohere, NousResearch, and Llama.
Hugging Face announced the acquisition of Seattle-based startup XetHub, founded by former members of Apple's ML infrastructure team, specializing in scalable Git-style data and model collaboration solutions for AI development.
Mistral AI announced the availability of customization capabilities for flagship and specialized models such as Mistral Large 2 and Codestral on La Plateforme, supporting base prompts.
OpenAI has launched a structured output feature for its API, allowing model outputs to strictly follow the JSON Schema provided by developers.
The official Hugging Face blog published an overview of Hub security features as of August 6, 2024, covering fine-grained tokens, two-factor authentication (2FA), and commit signing (GPG) for all users.
Hugging Face and Albumentations AI have collaborated to launch TextImage Augmentation, a multimodal data augmentation technique for document images. It supports random insertion, deletion, swapping, and stop-word replacement of text lines within images while preserving text semantics, with corresponding modifications to the image regions. The release includes complete installation commands, adaptation instructions for IDL/PDFA datasets, visualization examples, and usage guidelines for synthetic data generation.
DeepSeek API launches hard disk cache technology, reducing prices by an order of magnitude. See official documentation for details.
Google released Gemma 2 2B (a lightweight version with 2.6B parameters, supporting on-device deployment), ShieldGemma (an open-source safety classifier series based on Gemma 2, including 2B/9B/27B versions for input/output content moderation), and Gemma Scope (an open-source sparse autoencoder suite for Gemma 2 2B and 9B, used for model interpretability analysis).
A Hugging Face blog post introduces how to use the Quanto tool to apply quantization to Transformer-architecture diffusion models in the Diffusers library (such as PixArt-Sigma.
Hugging Face officially announced the launch of the NVIDIA NIM API (Serverless) inference service for Enterprise Hub organizations. It supports open-source large models such as Llama 3 and Mixtral, runs on NVIDIA DGX Cloud with H100 GPUs, uses an OpenAI-compatible API, and charges based on request duration; the service was discontinued on April 10, 2025.
PyTorch 2.4 has been officially released. Major updates include: torch.compile supports Python 3.12; AOTInductor adds a new freezing feature to support MKLDNN weight serialization and improve CPU performance.
Mistral AI released Mistral Large 2 (version 24.07, API name mistral-large-2407), featuring a 128k context window.
Meta has released the Llama 3.1 series of models, including three parameter scales: 8B, 70B, and 405B. All versions offer base and instruct-tuned variants and support a 128K token context.
Hugging Face published a complete solution in its blog for converting Mistral-7B-Instruct-v0.3 into a Core ML model runnable on Mac, based on new Core ML features from WWDC'24 (such as Swift Tensor, Stateful Buffers, and 4-bit block quantization).
Mistral AI and NVIDIA collaborated to release the Mistral NeMo 12B model, which supports a 128k token context window and utilizes FP8 quantization for lossless inference.
Hugging Face introduced the Multi-LoRA service feature in Text Generation Inference (TGI) v2.1.1, supporting the deployment of a single base model (e.g., mistralai/Mistral-7B-v0.1).
Hugging Face 发布 Docmatix 数据集，包含 2.4 百万张图像和 950 万组问答对，源自 130 万份 PDF 文档，规模是此前 DocVQA 数据集的 240 倍；使用 Phi-3-small 生成问答对并过滤 15% 幻觉样本；在 Florence-2 上微调后 DocVQA 性能提升 20%。
Mistral AI has released the Codestral Mamba model, based on the Mamba architecture with 7,285,403,648 parameters. It supports context retrieval of up to 256k tokens, is optimized for code and reasoning tasks, and can be tried on la Plateforme (codestral-mamba-2407). The original weights have been open-sourced on HuggingFace under the Apache 2.0 license.
Mistral AI has released Mathstral 7B, a math-specific large language model built upon Mistral 7B, achieving 56.6% on the MATH benchmark and 63.47% on MMLU.
The official Hugging Face blog published a tutorial demonstrating how to use distilabel to generate synthetic data and fine-tune the BGE base embedding model (plaguss/bge-base-argilla-sdk-matryoshka).
Hugging Face has released the SmolLM series of small language models, featuring three parameter sizes: 135M, 360M, and 1.7B. These models are based on the proprietary high-quality dataset SmolLM-Corpus (which includes Cosmopedia v2).
In collaboration with Hugging Face, Numina built the NuminaMath 7B TIR model based on DeepSeekMath-Base 7B, winning the first AI Mathematical Olympiad (AIMO) Progress Prize Kaggle competition with a score of 29/50.
Hugging Face has experimentally launched an automated PII detection reporting feature on Dataset Hub based on the open-source tool Presidio, used to identify sensitive personal information such as email addresses in datasets; this feature targets dataset users and owners, assisting in determining whether further filtering or verification of PII cleansing effectiveness is needed.
Hugging Face announced that the TRL library officially supports Direct Preference Optimization (DPO) for Vision-Language Models (VLMs). Using Idefics2-8b as an example, it demonstrates in detail the full workflow of DPO fine-tuning using the openbmb/RLAIF-V-Dataset.