Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Cohere For AI released the Aya Expanse series of models, including 8B and 32B parameter versions, utilizing data arbitrage and multilingual preference training.
Hugging Face and Google DeepMind released SynthID Text in Transformers v4.46.0, providing text watermark embedding and detection capabilities based on logit processors.
Hugging Face released CinePile 2.0, a long-video question-answering dataset enhanced through adversarial dataset refinement; this method uses a locally run LLaMA 3.1 70B as a 'deaf-blind LLM' to identify and rewrite weak Q&A pairs, successfully correcting 90.24% of degraded samples in the test set, while simultaneously open-sourcing all pipeline code and prompts.
Hugging Face officially released Generative AI Services (HUGS), a set of zero-configuration [services] built on Text Generation Inference and Transformers.
Hugging Face's official blog announced that Stable Diffusion 3.5 Large (8B parameters) and its timestep-distilled version are now available on Hugging Face Hub, with full support for 🧨 Diffusers. The model introduces architectural improvements such as QK normalization and dual attention layers, providing complete practical solutions including single-file loading, quantized inference (NF4), and LoRA fine-tuning, optimized for consumer GPUs.
Hugging Face and dottxt have collaborated to release outlines-core 0.1.0, a Rust implementation of the core algorithms for structured generation in outlines. It has been integrated into outlines, with Python bindings released simultaneously.
Hugging Face announced a partnership with Protect AI to integrate its Guardian scanning tool into the Hugging Face Hub, automatically detecting serialization formats in model files, such as pickle, that are vulnerable to arbitrary code execution attacks.
Hugging Face releases Transformers.js v3, with major updates including: WebGPU support (claimed to be up to 100 times faster than WASM), and new quantization formats (such as q4.
Hugging Face's official blog published a complete deployment guide for Speech-to-Speech (S2S) pipelines on Inference Endpoints, including building custom Docker images based on huggingface-inference-toolkit.
Hugging Face's official blog announced that Llama 3.2 models now natively support Keras Hub, allowing direct loading from Hugging Face via keras_hub.models.Llama3CausalLM.from_preset.
PyTorch 2.5.0 has been officially released. Core updates include: adding a default-enabled CuDNN backend for scaled_dot_product_attention, delivering up to 75% speedup on H100 and newer GPUs.
Mistral AI has released two edge computing models, Ministral 3B and Ministral 8B, supporting up to 128k context (currently vLLM supports 32k).
The official Hugging Face blog disclosed an issue where gradient accumulation in the transformers library caused incorrect loss calculations in token-level tasks such as causal language modeling: the original implementation averaged the loss per batch.
Hugging Face's official blog published an AI inference performance validation for AMD's fifth-generation EPYC CPUs (Zen5 architecture, codenamed Turin), featuring measured results for Llama 3.1 8B in Summarization.
Hugging Face officially released Gradio 5 after commissioning Trail of Bits to conduct a comprehensive security audit of the codebase. The audit covered four scenarios: local execution, Spaces deployment, built-in share links, and CI supply chain. All identified security issues, including SSRF, RCE, CORS misconfiguration, and file upload vulnerabilities, were found and fixed, with all patches integrated into version Gradio 5.0.
Hugging Face's official blog introduces how to extend AI data processing capabilities using Dask, demonstrating the process starting from local 100-row pandas handling with the FineWeb-Edu classifier as an example.
Hugging Face officially announced the stable release of Gradio 5, featuring server-side rendering for faster loading, modern UI components with built-in themes, and low-latency streaming (including WebRTC and audio/video use cases).
Hugging Face, in collaboration with Intel Labs, introduced Dynamic Speculative Decoding, which serves as the default execution mode for assisted generation in Transformers 4.45.0.
Addressing storage efficiency issues for over 2.2PB of Parquet data on the Hub, the Hugging Face Xet team quantified byte-level deduplication effects by performing append, modify, and delete operations on 2GB Parquet files from FineWeb. They discovered that absolute offsets causing column header rewrites are the primary reason for reduced deduplication rates in modification/deletion scenarios, and proposed an experimental improvement called Content Defined Row Groups.
The Hugging Face official blog announced the launch of the Open FinLLM Leaderboard, an open-source large language model evaluation platform designed specifically for the financial domain.
OpenAI has released the Realtime API, which developers can integrate into their applications to build fast voice-to-voice experiences.
Developers can now fine-tune GPT-4o using image and text data to improve its visual capabilities.
OpenAI has banned STORM-2035 accounts originating from Iran, which used AI to generate content related to US and UK elections and published it across multiple websites.
OpenAI has banned 'Stop News' accounts originating from Russia, which used AI to generate multilingual articles and posts for influence operations targeting Ukraine and Western countries.