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 and AMD have released open-source kernels optimized specifically for the MI300X GPU, including fused residual connection/RMS norm/FP8 conversion and fused SwiGLU activation/FP8 conversion.
Hugging Face's official blog announced that Gradio supports the Model Context Protocol (MCP) starting from version 5.28.0, enabling Hugging Face Spaces built with Gradio to serve as MCP servers for LLM clients (such as Cursor.
Pollen Robotics and Hugging Face jointly released the open-source robot Reachy Mini, offering Lite and Compute versions priced at $399 and $499 respectively. It supports Python programming, simulation development, and Hugging Face model integration, and is currently in the early development stage.
Based on the nanoVLM project, the Hugging Face team discovered that traditional multimodal training suffers from severe GPU idling and padding waste in data pipelines, proposing a five-stage progressive optimization solution.
Hugging Face releases the open-source SmolLM3-3B model, supporting a 128k context window, 6 languages, and dual-mode inference (/think and /no_think), outperforming Llama-3.2-3B and Qwen2.5-3B at the 3B scale.
The official Hugging Face blog details how to train and fine-tune sparse embedding models (such as SPLADE and CSR architectures) using the Sentence Transformers library.
Sakana AI has released AB-MCTS (Adaptive Branching Monte Carlo Tree Search) and Multi-LLM AB-MCTS inference-time scaling algorithms.
NVIDIA has released the 8B vision-language model Llama Nemotron Nano VL, designed specifically for intelligent document processing. It supports high-precision text recognition, table and chart parsing, element localization (bounding box), and multi-format output (HTML/LaTeX/markdown).
Google released Gemma 3n, a multimodal model natively supporting image, text, audio, and video inputs, now open-sourced on Hugging Face. Two sizes were released: gemma-3n-E2B (5B actual parameters, requiring only 2GB GPU memory) and gemma-3n-E4B (8B actual parameters, requiring only 3GB GPU memory), both including base and instruct versions.
SGLang has added backend support for the Hugging Face transformers library, enabling direct execution of any transformers-compatible model (such as meta-llama/Llama-3.2-1B-Instruct or kyutai/helium-1-preview-2b).
The official Hugging Face blog published a tutorial demonstrating the use of the diffusers library and QLoRA technology to fine-tune the black-forest-labs/FLUX.1-dev model on a single NVIDIA RTX 4090 (24GB VRAM).
Hugging Face officially announced that Groq has become a supported inference service provider on its Hub platform, supporting open-source models such as Llama 4 and QWQ-32B; Groq provides low-latency, high-throughput inference based on LPU™ technology, accessible via custom API Keys or routing through Hugging Face accounts; billing is settled according to the actual provider or HF account, with PRO users receiving $2 in monthly inference credits.
TNG self-hosts multiple large language models on 24 H100 GPUs, processing over 100 million input tokens daily; the article points out that under vLLM's default chunked-prefill strategy.
Mistral AI announced the launch of Mistral Compute, a private AI infrastructure service for sovereign nations, enterprises, and research institutions. It provides GPUs, orchestration, APIs, and accompanying tools ranging from bare-metal servers to fully managed PaaS. The initial deployment supports tens of thousands of NVIDIA GPUs, emphasizing data sovereignty, sustainability, and European compliance. Black Forest Labs, BNP Paribas, and several other organizations have joined as founding partners.
Mistral AI has released its first reasoning model, Magistral, which includes an open-source version, Magistral Small (24B parameters), and an enterprise version, Magistral Medium.
Figure officially announced that its Helix logistics agent achieved significant upgrades within three months. It can now stably handle non-rigid packages such as deformed polyethylene bags and flat envelopes. The average processing time per package has dropped to 4.05 seconds, and the barcode orientation success rate has improved to approximately 95%. Improvements include a visual memory module, state history input, and force feedback integration, trained on 60 hours of demonstration data.
Mistral has released Mistral Code, an AI coding assistant for enterprises that integrates four models: Codestral, Codestral Embed, Devstral, and Mistral Medium. It supports private beta for JetBrains IDEs and VSCode, offering local/cloud/air-gapped deployment, fine-grained RBAC, audit logs, and fine-tuning capabilities for private repositories.
OpenAI banned accounts associated with the STORM-2035 operation, which is potentially backed by Iran and uses AI to generate influence content related to politics in the United States, the United Kingdom, Ireland, and Venezuela.
Sakana AI, in collaboration with Jeff Clune's lab at the University of British Columbia, proposed the Darwin Gödel Machine (DGM), a self-improving agent that continuously enhances its programming capabilities by rewriting its own Python code.
Mistral AI released Codestral Embed, the first embedding model dedicated to code, supporting different dimensions (e.g., 256) and precisions (e.g., int8), on SWE-Bench lite.
DeepSeek-R1 has been upgraded to DeepSeek-R1-0528, featuring enhanced reasoning capabilities with significant Pass@1 score improvements on benchmarks including AIME 2025, GPQA, LCB_v6, and Aider. Web front-end development capabilities have been optimized, resulting in more aesthetically pleasing generated web pages and games. Hallucinations have been significantly reduced. The model now supports JSON Output and Function Calling, achieving Tau-bench scores of Airline 53.5 / Retail 63.9.
Mistral has released the Agents API, offering built-in connectors for code execution, image generation (based on FLUX1.1 [pro] Ultra), document libraries (supporting RAG), and web search. It supports MCP tool integration, stateful conversations, and multi-agent orchestration.
OpenAI has launched Stargate UAE, the first international deployment version of its AI infrastructure platform, Stargate.
Mistral AI, in collaboration with All Hands AI, has launched Devstral, an agentic large language model designed for software engineering tasks, achieving a score of 46.8% on the SWE-Bench Verified benchmark.