Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
Zhipu released the GLM-5.2 large language model, achieving stable and usable 1M token context in an open-source model for the first time, specifically designed for long-horizon tasks; the new model adopts the IndexShare architecture to reduce FLOPs and improves the MTP layer to increase speculative decoding acceptance length by 20%.
Hugging Face, in collaboration with Microsoft, Google, and others, released the open Agentic Resource Discovery (ARD) specification, defining a mechanism for dynamic discovery of agent capabilities.
The Google Open Source Blog announced the launch of the Universal Commerce Protocol (UCP) in collaboration with industry partners. UCP is an open protocol for agentic commerce, designed to provide a unified interaction language for consumer interfaces, merchant systems, and payment service providers. The protocol features a layered architecture covering service domains, core capabilities (such as cart, checkout, and orders), configurable extensions, and multiple transport bindings (REST/MCP/Agent2Agent). The protocol is open source, has established a technical committee, and continues to evolve.
Google DeepMind, in collaboration with the UK government, Google Cloud, Faculty, and local planning authorities in Barnet, Dorset, and Camden, has developed an AI planning approval prototype tool based on Gemini. The tool aims to reduce the approval time for residential planning applications by 50%. It supports data integration, policy identification, feedback summarization, and generation of draft reports. Currently piloted in three locations, it is planned for rollout to councils across the UK in 2027.
Google Research released the Earth AI framework, utilizing high-resolution satellite imagery, 1-meter LiDAR, and pre-trained Vision-Transformer (ViT) models.
Zhipu has launched and open-sourced GLM-5.2. The model scored 51 points on the Artificial Analysis comprehensive leaderboard, achieving state-of-the-art (SOTA) status among open-source models; it ranked first globally among available models in the Code Arena blind test.
Google DeepMind released the "AI Control Roadmap," proposing a system-level security framework for internally deployed AI agents. The core concept is to treat unaligned agents as 'insider threats,' combining this with MITRE ATT&CK to build a threat classification taxonomy. It supports actual deployments such as Gemini Spark through supervised AI real-time monitoring, tiered responses (Detection levels D1–D4, Response levels R1–R3), and analysis of millions of coding agent trajectories. Simultaneously, it released a technical report titled "Three-Layer Framework for Agent Safety" aimed at policymakers.
The Tongyi team released the Qwen-Robot Suite, which includes Qwen-RobotNav (unifying five types of navigation tasks, trained on 15.6 million samples, supporting zero-shot real-robot deployment), Qwen-RobotManip (based on >38.
QwenTeam released Qwen-RobotManip, a general vision-language-action (VLA) foundation model built on Qwen-VL, proposing a unified representation, motion, and behavior three-dimensional cross-embodiment alignment framework.
The Qwen Team released the Qwen-RobotWorld technical report, proposing an embodied world model that uses natural language as a unified action interface and adopts a Dual-Stream Multimodal Diffusion Transformer (MMDiT) architecture.
The QwenTeam released Qwen-RobotNav, a scalable navigation model built on Qwen3-VL, trained on 15.6 million samples, with a single set of weights unifying VLN and ObjNav.
vLLM v0.23.0 has been officially released. Core updates include the maturation of DeepSeek-V4 across multiple backends (decoupling MLA metadata, adding TRTLLM-gen attention kernels, EPLB support, etc.).
SGLang v0.5.13 has been officially released, adding support for autoregressive models including Nemotron 3 Ultra, Step-3.7-Flash, Command A+, and Cosmos3.
The Google Go team has released the pkg.go.dev v1beta API, providing structured JSON interfaces for querying metadata such as Go modules, packages, symbols, versions, vulnerabilities, and dependencies. A reference CLI tool, pkgsite-cli, was also released alongside an open OpenAPI specification. This API supports toolchain construction and AI-assisted coding scenarios.
Google Research published a blog post introducing the 'Phone Cluster Computing' research conducted in collaboration with the University of California, San Diego: extracting motherboards from retired Pixel phones, removing non-computing components, replacing the Android user space with a general-purpose Linux distribution, and using Kubernetes to orchestrate 25–50 phones into a self-managing cluster; plans are in place to deploy a cluster composed of 2,000 phones in Fall 2026, providing computing power equivalent to approximately 50 servers to support teaching and research applications.
Google GKE Labs has open-sourced OpenRL, a self-hosted post-training API deployable on proprietary Kubernetes clusters that supports LLM fine-tuning. It abstracts RL training infrastructure, provides four core API types, and is compatible with the Tinker ecosystem. Currently focused on LoRA fine-tuning, it has released a text-to-SQL example (based on Gemma models) and quick-start guides for local, Mac, and GKE environments.
Data from Marketplace Pulse based on nearly 100,000 US TikTok Shop sellers shows that the top 1% of sellers contribute 60% of cumulative GMV, with the top 0.1% (fewer than 90 stores) contributing over a quarter; this concentration is higher than that of Amazon, and platform certification badges (Official Store, Gold Star) further amplify the gap in traffic and sales.
Starting in April 2026, Google became a strategic member of the Eclipse Foundation, sponsoring Open VSX and being among the first to adopt its hosted registry service; Open VSX is an open-source, vendor-neutral extension registry based on the VS Code™ Extension API, supporting AI-integrated IDEs and development platforms such as Google Antigravity, AWS Kiro, Cursor, and Windsurf; currently, Open VSX sees over 300 million downloads per month, daily peak requests reaching 200 million, hosting over 12,000 extensions from more than 8,000 publishers.
Google Research proposes the Regularized f-Divergence Kernel Tests framework for more sensitive, flexible, and accurate auditing of machine unlearning; this framework replaces traditional two-sample tests with three-sample relative tests.
Google DeepMind has released the experimental open-source text diffusion model DiffusionGemma, featuring a 26B MoE architecture that activates only 3.8B parameters during inference and supports 256-token parallel generation.
Google DeepMind has released Gemini 3.5 Live Translate, a near-real-time speech-to-speech translation audio model supporting over 70 languages that preserves tone, rhythm, and pitch.
Google DeepMind has released Gemma 4 12B, a unified multimodal large language model that does not require visual or audio encoders. It supports native image and audio inputs and can run locally on consumer-grade laptops equipped with 16GB of VRAM.
In its blog post, MiniMax detailed the technical path behind the leap in its mathematical reasoning capabilities, centered on the MaxProof framework: employing Proof RL (using a layered defense reward system with multi-model and multi-mode generative verifiers).
The author uses a coding agent to chain two Hugging Face Spaces: ideogram-ai/ideogram4 (image generation) and VAST-AI/TripoSplat (generating 3D Gaussian splats from a single image).