Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
NVIDIA released the open-source 30B MoE model Nemotron 3.5 Lightning, which activates only 3B parameters and is specifically designed for the high-throughput, low-latency execution layer (such as tool calling) of long-running AI agents.
NVIDIA has released NeMo Switchyard, an open-source model routing system that supports dynamically dispatching requests to different models (such as DeepSeek V4 and Kimi K2.6) based on tasks within agent workflows.
Mistral AI announced the launch of Mistral Regional Endpoints (generally available, supporting inference in European or US regions), Mistral Priority Tier (in public beta, including custom rate limits and SLA guarantees), and began supporting third-party open models on its infrastructure (starting with Z.ai's GLM-5.2). The company also launched the European Compute Unit (ECU) initiative, aiming to build up to 1 GW of European AI compute infrastructure through long-term enterprise commitments.
OpenAI has begun testing ads in ChatGPT to support free access, while ensuring that ads are clearly labeled, do not affect response content, include strong privacy protections, and provide users with control.
Ollama v0.32.8 has been released, officially supporting the Muse Glimmer model on all platforms including Apple Silicon, NVIDIA, and AMD. It also provides dedicated launch commands for agent frameworks such as Claude Code, Pi, OpenClaw, and Hermes; the Ollama MLX engine now supports DFlash and image input.
vLLM v0.27.0 has been officially released, including 561 commits, supporting full-stack implementation of Kimi K3 (including core model files, Python/Rust frontends, AttnRes kernels, DeepGEMM.
GitHub has released the Copilot SDK for Java, a framework-agnostic, AI-vendor-neutral client library that supports any OpenAI-compatible endpoint, including OpenAI, Azure, and Anthropic.
NVIDIA has released the Magpie multilingual TTS model with open-source weights, supporting a total of 12 languages: English, Spanish, French, German, Italian, Vietnamese, Mandarin Chinese, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese.
Meta has released Muse Glimmer, a 30-billion-parameter open-source dense architecture large language model supporting an ultra-long context window of over 120K tokens. It is designed specifically for local AI agent workflows and achieves a throughput of 20K tokens/sec on a single NVIDIA GPU.
Ollama v0.32.7 has been released, marking the first time it supports the 30B multimodal model Muse Glimmer from Meta Superintelligence Labs on the MLX engine. Currently limited to Apple Silicon, it supports DFlash and image input; `ollama run` and `ollama launch` commands are provided for use in local Agent applications such as Claude Code, Codex, Pi, OpenClaw, and Hermes.
Hugging Face and Multiverse Computing jointly released the paper "Efficient Knowledge Distillation for LLMs."
Meta released the Muse Glimmer-30B multimodal large language model, distilled from Muse, with 30B parameters and under the Apache 2.0 open-source license; it supports joint reasoning over images, video (without audio), and text.
SGLang v0.5.17 has been officially released, adding day-0 native support for Kimi K3 (a 2.8T-parameter LatentMoE multimodal model with a 1M-token context) and MiniMax-H3 (a text/frame/reference-driven video + audio generation model).
Monthly spending on crypto payment cards reached $759 million in July 2026, an increase of approximately 2.5 times compared to one year prior; USDC accounted for 58% and USDT for 26%, while EURe's share dropped from 88% in early 2024 to 2%; Optimism, Solana, and Base each handled approximately 19%-29% of settlement volume, with Gnosis falling to about 2%; data is sourced from Paymentscan, while RedotPay spending figures are self-reported by the issuer.
Google DeepMind released and open-sourced the WeatherNext series of AI meteorological models, including WeatherNext 2, WeatherNext Cyclones, and the lightweight version WeatherNext 2-mini.
Hugging Face announced that Baseten has officially become an inference provider supported by its Hub platform. The initial support covers chat and text generation tasks, including Kimi K3 and DeepSeek-V4-Flash-0731.
a16z crypto released an in-depth Q&A analyzing how the "CLARITY Act" (HR 3633) establishes a federal regulatory framework for the US digital asset market: it clarifies the phased regulatory responsibilities of the SEC and CFTC over digital assets (SEC during the centralized phase, transitioning to CFTC after decentralization), prohibits interest on stablecoin balances while allowing transaction rewards, exempts general open-source software developers from unlimited liability for unforeseen downstream uses, and strengthens exchange customer asset protection and anti-fraud requirements. The act aims to end the 'race to the bottom' caused by regulatory ambiguity and prevent a recurrence of the FTX incident.
a16z crypto published a long-form analysis of the "CLARITY Act" (HR 3633), which aims to establish a federal regulatory framework for the U.S. digital asset market. The bill clarifies the jurisdictional division between the SEC and CFTC over digital assets at different stages, prohibits interest on stablecoin balances while allowing transaction rewards, exempts general open-source software developers from unlimited liability for unforeseen downstream uses, and strengthens requirements for exchange customer asset protection and anti-fraud measures.
AMD has launched the AUP Learning Cloud, an AI teaching platform based on Kubernetes and JupyterHub, optimized for AMD ROCm hardware, supporting deployments ranging from a single AI PC to campus-level GPU clusters.
Tencent Hunyuan Hy3 has officially opened globally, supporting access via WorkBuddy (free until August 31, 2026), Miora, Tencent Cloud TokenHub, and API; it adopts a MoE architecture integrating fast and slow thinking.
The Ethereum Foundation is providing funding to the Freedom of the Press Foundation (FPF) through the Trillion Dollar Security (1TS) program to support the ongoing development of the open-source tool WEBCAT. WEBCAT enables browsers to verify that code served by registered websites matches the code published by their developers; it is currently implemented as a Firefox extension. This grant will drive the development of a verification library integrable with wallets, adaptation to Chromium-based browsers such as Chrome, independent security audits, and the establishment of ERC standards.
The GitHub engineering team introduces how to use Stacked Pull Requests (stacked PRs) to decompose massive PRs generated in one go by AI agents (e.g., 1721 lines of changes) into four small PR layers: L1–L4, covering data models, search APIs, chat calls, and UI with citations; each layer is independently reviewable with clear dependencies, supporting management via the gh stack CLI, automatic rebasing, and syncing.
Doubao officially launched the native audio-video full-duplex large model SeedRealtime. It adopts a unified end-to-end architecture to natively integrate audio, video, and text, enabling real-time interaction on continuous multimodal streams. The model supports joint audio-video understanding, proactive interaction, and smooth conversational rhythm. This model has been fully released in the Doubao App; users can experience it by updating to the latest version and selecting "Make a Call" in the video call interface.
NVIDIA's technical blog published research on World Action Models (WAM), noting that WAM uses video world models as the backbone to replace traditional Vision-Language-Action models (VLA), enabling better modeling of physical dynamics and improving open-world generalization.