Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
Ollama v0.32.11 has been released, adding support for DeepSeek Harness via 'ollama launch dsh' and Meta Muse Code via 'ollama launch muse'. The OpenAI-compatible response API now includes web search capabilities, and the Muse Glimmer template has been updated. See the GitHub link for the full changelog.
Hugging Face released the "State of Open Source Models Summer 2026" report. Based on Hub data from January to August 2026, it noted that Chinese labs skipped the small-model stage and directly released models with parameter counts ranging from hundreds of billions to 2.78 trillion.
Cursor announced that the Firetiger team has joined. The agent built by Firetiger can monitor releases, detect regressions, investigate incidents, and feed issues back to coding agents after software goes live; this capability will be integrated into various Cursor features, working with Cursor Origin and the upcoming Change Monitors to connect the agentic workflow from code writing to production runtime.
Hugging Face launched Storage Buckets (released in March 2026, backed by Xet, non-versioned, hf:// namespace), supporting Strands Robots (an AWS open-source SDK).
Google DeepMind released Gemini 3.7 Flash, claiming it is currently the most intelligent model for coding and agentic workloads; compared to Gemini 3.6 Flash, it achieved 43.6% vs 34.4% on FrontierCode 1.1 Main.
MiniMax officially released Music 3.0, a new generation open-weight music generation model that supports generating complete songs (up to 5 minutes) in a single pass. Core upgrades include: a creative intent understanding framework based on fine-grained temporal descriptions, a Hybrid-LM architecture with global-local coordination (8B Global LLM + 0.6B Local LLM), and Flow-VAE audio synthesis driven by multi-layer RVQ encoding and continuous latent state fusion.
Cursor announced it has obtained AIUC-1 certification, a standard supported by researchers from MITRE, Cloud Security Alliance, and Stanford, which integrates independent audits with thousands of adversarial scenario tests.
Cursor has launched a build feature that improves cloud agent startup speed by up to 3x by pre-preparing copies of development environments in the background. Builds are created hourly by default; starting August 17, this feature will be enabled by default for all new and existing environments with no additional charges.
OpenAI has launched Ultrafast mode, a new API service tier running GPT-5.6 Sol, with inference speeds up to 14 times faster, powered by Cerebras, achieving output throughput of up to 750 tokens/second.
From July 15 to August 2, 2026, Hugging Face launched the ICML 2026 Open Reproducibility Challenge, in which 1,221 community members used Claude Code and Codex.
Alibaba has released the open-source large language model Qwen3.8-2.4T-A95B (Qwen3.8-Max), featuring a total parameter count of 2.4T with 95B activated per token, utilizing a fine-grained MoE architecture and a hybrid design of full attention and linear attention.
AllenAI has launched a custom export feature for embedding vectors on the OlmoEarth Studio platform, enabling users to filter by region, time range, and encoder variant (Nano/Tiny/Base).
NVIDIA's technical blog published a full-stack observability practice guide for AI factories. It proposes a decision framework covering fault domains such as platform health, GPUs, InfiniBand networks, and cluster jobs; clarifies the responsibility boundaries and combination logic of tools including DCGM, NVSM, UFM, and BCM; and provides methods for tool selection in InfiniBand clusters, defining Top-K alert sets, and building unified dashboards using Prometheus/Grafana.
Microsoft Research released the MindTopo benchmark to evaluate the reasoning and planning capabilities of multimodal large language models (VLMs) across five types of topological relationships: connectivity, separation, order, closure, and knotting. Results show that existing models perform adequately on static recognition tasks but fail significantly in interactive planning tasks requiring maintenance of topological relationships across actions, with errors occurring primarily during the planning stage rather than the perception stage.
The official version of DeepSeek-V4-Pro has been launched on the APP, web interface, and API, with the model name deepseek-v4-pro; Agent capabilities have significantly improved, achieving an HLE score of 42.7/60.0.
Google DeepMind 发布 SL2T 手语转文本（SL2T）模型，首次将手语 AI 集成至消费级产品：在 Pixel 11 上通过 Gboard 和 Live Transcribe 实现美国手语（ASL）到英文的实时转写。
Google Research published the paper "Empty shelves or lost keys? Recall is the bottleneck for parametric factuality".
Cursor and SpaceXAI jointly released Grok 4.6, a model focused on long-running agents, complex interactions, and visual work, which is now available in Cursor and Grok Build.
NVIDIA JetPack 7.2.1 has been officially released, introducing PyNvVideoCodec 2.2 (a hardware-accelerated video codec library with Python interfaces) for the first time. Based on this, Jetson agentic video skills have been built, supporting device discovery, encoding configuration generation, performance benchmarking, and workflow validation. Additionally, Jetson T5000 now offers performance simulation capabilities for T3000.
Google announced its formal joining of the OpenROAD Initiative (ORI) as a principal member. Aaron Cunningham will represent Google on the ORI Board. ORI is a non-profit organization dedicated to the open-source electronic design automation (EDA) ecosystem. Its core project, OpenROAD, is an open-source, fully automated digital chip design toolchain that supports the RTL-to-GDSII flow within 24 hours. Google's membership aims to support neutral governance, ecosystem expansion, talent development, and technical enhancement, with backing from the NSF POSE program.
Google Research released AMIE (Video), a real-time video consultation AI system built on Gemini and Project Astra, capable of perceiving non-verbal clinical cues during synchronous video sessions.
Microsoft Research has released the research model CARE-X, aimed at enhancing the clinical utility of radiology AI. It is based on the SigLIP2 visual encoder and the Phi-4-mini-instruct language model, integrating generative and discriminative capabilities.
IBM Research and Hugging Face propose ALTK-Evolve, an agent memory method that requires no weight updates and does not compress experience. It reduces inference token consumption by selectively retrieving guidelines per task (rather than injecting them in full).
Ollama v0.32.9 has been released, adding support for the NVIDIA Nemotron 3.5 Lightning open-source MoE model (30B total parameters, 3B active parameters), designed specifically for resident agent execution layers, compatible with OpenClaw and Hermes Agent; it also adds support for the Nemotron 3 architecture and fixes boundary condition issues in Muse Glimmer's function call parsing.