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 announced that DeepInfra has officially joined its Inference Providers ecosystem, supporting calls to models hosted by DeepInfra via model pages, SDKs (Python/JS), and Agent Harnesses.
NVIDIA has released Nemotron 3 Nano Omni, a unified understanding model supporting long-context text, image, video, and audio inputs. It features capabilities in document analysis, ASR (Automatic Speech Recognition), joint audio-video understanding, agentic computer operation, and general multimodal reasoning.
AI21 发布 Maestro 智能体优化框架，用于系统化搜索智能体在质量、成本与延迟三维度间的最优权衡点；在 BrowseComp-Plus 上实现 95.18% 准确率的 SOTA 性能。
OpenAI GPT models, Codex, and managed agents are now available on AWS, enabling enterprises to securely build AI within their own AWS environments.
vLLM v0.20.0 has been officially released, adding support for multimodal and cutting-edge models including DeepSeek V4, the preview version of Hunyuan v3 (Hy3), and Granite 4.1 Vision.
OpenAI has obtained FedRAMP Moderate authorization for ChatGPT Enterprise and the OpenAI API, supporting the secure adoption of AI by U.S. federal agencies.
Mistral AI has released a public preview of Workflows, an AI orchestration layer built into Studio that supports durable execution, observability, human-in-the-loop approvals, and RBAC for enterprise capabilities. Built on Temporal, the control plane is hosted by Mistral while worker nodes are deployed in customer environments; Python SDK v3.0 is now available for installation.
OpenAI has open-sourced Symphony, an open specification for Codex orchestration that transforms issue tracking systems into always-on agent systems, enhancing engineering output and reducing context switching.
OpenAI has released the open-source PII detection model Privacy Filter on Hugging Face Hub, supporting a 128k context window and identifying eight categories of sensitive information in a single forward pass.
Sakana AI released the paper "Learning to Orchestrate Agents in Natural Language with the Conductor," which was accepted by ICLR2026.
Sakana AI publishes the paper "TRINITY: An Evolved LLM Coordinator," proposing a test-time model composition system: The TRINITY coordinator does not modify underlying model weights and uses a lightweight routing head (<20K learnable parameters) in Thinker.
DeepSeek has released the V4 model series, including DeepSeek-V4-Pro (1.6T total parameters / 49B active) and DeepSeek-V4-Flash (284B total parameters / 13B active), both supporting a 1M-token context window.
The DeepSeek API now supports the V4-Pro and V4-Flash models, compatible with OpenAI ChatCompletions and Anthropic interfaces.
Hugging Face blog published a tutorial detailing how to build a Chrome browser extension based on Transformers.js and Gemma 4 E2B, covering Manifest V3 architecture design (background service worker.
Google Photos has introduced the Auto Frame feature, which uses an internal 3D point map estimation model and a generative latent variable diffusion model to parse photos into 3D scenes and automatically adjust virtual camera pose and focal length, thereby reconstructing portrait perspectives. This feature is now live, allowing users to select the second version of automatic reconstruction results for eligible portrait photos.
Google DeepMind released a new paper proposing the Decoupled DiLoCo (Distributed Low Communication) architecture, which divides training into computational 'islands' (learner units) with asynchronous communication.
ChatGPT has launched workspace agents based on Codex, which can automate complex workflows in the cloud, helping teams securely scale work across tools.
The Tongyi Qianwen team has open-sourced Qwen3.6-27B, a 27 billion parameter dense multimodal model, on agentic coding benchmarks (SWE-bench Verified, Terminal-Bench 2.0).
OpenAI has released the OpenAI Privacy Filter model with open weights, designed to detect and anonymize Personally Identifiable Information (PII) in text, achieving state-of-the-art accuracy.
In an ICLR paper, Google Research proposed ReasoningBank, a novel agent memory framework that extracts structured reasoning patterns from both successful and failed experiences, supporting self-evolution at test time.
Hugging Face, in collaboration with TII, released QIMMA قِمّة, a quality-first Arabic LLM evaluation benchmark. It conducted multi-stage validation (dual LLM automatic scoring + native speaker manual review) on over 52,000 samples from 14 existing benchmarks, discovering systematic quality issues ranging from 0.2% to 3.1% in benchmarks such as ArabicMMLU. QIMMA integrates 7 major domains, includes Arabic code evaluation for the first time, and publicly releases all sample-level inference outputs.
Moonshot AI has open-sourced Kimi K2.6, supporting Kimi.ai, the Kimi App, API access, and Kimi Code. The model significantly outperforms K2.5 in long-horizon coding tasks; for example, it deployed the Qwen3.5-0.8B model using Zig on a local Mac with over 4,000 consecutive tool calls executed within 12 hours, and autonomously refactored the exchange-core engine, resulting in an 185% improvement in medium throughput.
Moonshot AI has open-sourced its latest model, Kimi K2.6, which supports long-horizon coding, Agent Swarm, and proactive agents (such as OpenClaw and Hermes). It is now available on Kimi.com, the Kimi App, via API, and in Kimi Code.
The Tongyi Qianwen team released the Qwen3.6-Max-Preview version, which outperforms Qwen3.6-Plus in agentic programming (SkillsBench +9.9, SciCode +6.3, etc.).