Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
SGLang has released version v0.5.10, with major updates including: Piecewise CUDA graphs enabled by default; integration of Elastic NIXL-EP to achieve partial fault tolerance for DeepSeek MoE.
AI21 notes that the bottleneck in AI system development has shifted from translating natural language into code to converting human subconsciousness into explicit specifications. It analyzes the 'I know what it is' paradox, the Dwight fallacy, speech act traps, and the limitations of Claude Code. It proposes a four-stage Active Discovery method: source mapping, simulation loops, blind spot detection, and coverage design, emphasizing that PRDs should be dynamic configurations rather than static documents.
Arvid Kahl notes that while publicly sharing revenue data and product details previously attracted customers and acquirers, CLI agentic coding tools now allow others to quickly replicate products using only prompts. He advises founders to stop sharing information easily used as blueprints—such as financial figures, system architecture, and tech stacks—and instead share generalized industry insights and operational lessons, with the core principle being "engagingly interesting, hard to clone."
vLLM has released version v0.19.0, adding full architecture support for Gemma 4 (including MoE, multimodal, reasoning, and tool calling), zero-bubble asynchronous scheduling with speculative decoding, and maturation of Model Runner V2.
Google DeepMind has released the Gemma 4 series of open-source large language models, featuring four specifications: Effective 2B, Effective 4B, 26B MoE, and 31B Dense.
Google DeepMind has launched the Gemma 4 multimodal model family, comprising five sizes: E2B, E4B, 12B Unified, 26B A4B, and 31B. All models are released under the Apache 2.0 open-source license, support joint input of images, audio, video, and text, and offer long-context text output (up to 256K).
The Tongyi Qianwen team officially released the Qwen3.6-Plus model, available via the Alibaba Cloud Bailian API. It supports a default 1 million token context window, significantly enhances agentic programming, multimodal perception, and reasoning capabilities, and introduces the preserve_thinking parameter to optimize reasoning consistency for long-horizon tasks. The model has been adapted for mainstream coding agent tools such as OpenClaw, Claude Code, and Qwen Code.
Zhipu has released GLM-5V-Turbo, a multimodal coding foundation model designed for visual programming. It natively supports multimodal inputs including images, videos, and design mockups, features a context window of up to 200k, and possesses multimodal tool-calling capabilities such as drawing bounding boxes, taking screenshots, and reading web pages.
TII Falcon Vision Team released Falcon Perception on the Hugging Face blog, a 0.6B parameter early-fusion Transformer model.
Playwright v1.59.0 has been released, featuring a new page.screencast API that supports page recording with action annotations, chapter titles, real-time frame streaming, and visual overlays, which can be used to generate video credentials for AI coding agents. A new browser.bind() method enables cross-process browser sharing, allowing connections from playwright-cli, @playwright/mcp, and other clients. A new playwright-cli show command opens an observability Dashboard. Enhanced CLI debugging and trace analysis capabilities support automatic test repair by agents. The await using syntax is introduced to support automatic cleanup of asynchronous resources. Locator and snapshot capabilities are enhanced with additions like ariaSnapshot and pickLocator. Breaking changes include the removal of macOS 14 WebKit support.
Hugging Face officially released gradio.Server in a blog post. It is a backend service framework based on FastAPI extensions, integrating Gradio queue, concurrency control, SSE streaming responses, ZeroGPU support, and gradio_client compatibility.
Google Research published a paper exploring the trade-off between the number of human raters (K) and the evaluation sample size (N) in AI benchmarks, noting that the common practice of using 1–5 raters per sample is often insufficient to capture human disagreement.
IBM has released Granite 4.0 3B Vision, a LoRA adapter-based vision-language model built on Granite 4.0 Micro, specifically designed for complex enterprise document understanding.
Mistral AI has released Spaces CLI, a command-line tool supporting collaborative use by human developers and coding agents; it achieves this by providing flag equivalents for all interactive inputs and generating context.json and AGENTS.md files.
OpenMed released the CodonRoBERTa series of mRNA language models on the Hugging Face blog, including the single-species CodonRoBERTa-large-v2 (trained on E. coli data.
Google Research released a white paper stating that future quantum computers may be able to break elliptic curve cryptography based on ECDLP-256 in minutes using fewer than 500,000 physical qubits. The paper proposes two implementations of Shor's algorithm circuits (requiring <1200/<1450 logical qubits, respectively) and employs zero-knowledge proofs to validate the conclusions without exposing the attack path. It also calls for migration to post-quantum cryptography (PQC) and provides short-term and long-term security recommendations for blockchain.
Hugging Face officially released TRL v1.0, upgrading the original research codebase into a production-grade post-training library with clear stability contracts, supporting over 75 methods. It adopts a design where the stable core (SFT, DPO, GRPO, etc.) coexists with the experimental layer (e.g., ORPOTrainer), addressing frequent paradigm shifts in the field through minimal abstraction, localized explicit implementations, and tolerance for controlled duplication. It provides a migration guide, emphasizing that the upgrade cost from version 0.x is extremely low.
Sequoia Capital investors Jack Dorsey and Roelof Botha published a long-form article analyzing the evolution of organizational structures from the Roman legions and the Prussian General Staff to modern enterprises. They argue that for two thousand years, hierarchical structures have essentially been 'routing protocols' constrained by human information processing capabilities. Block is building a new type of company form based on a Company World Model and a Customer World Model. Through an intelligence layer, it dynamically combines atomic capabilities (such as payments and lending) and delivers solutions via interfaces. Organizational roles are simplified to Individual Contributors, Directly Responsible Individuals (DRIs), and Player-Coaches, with the goal of using AI systems to replace humans as the coordination mechanism.
QwenTeam officially released Qwen3.5-Omni, a new generation native full-modal large model supporting understanding and generation of text, images, audio, and audio-video; it adopts the Thinker-Talker Hybrid-Attention MoE architecture.
Google DeepMind has released an AI pointer feature based on Gemini. It is currently integrated into Chrome, allowing users to point directly at web content to ask Gemini questions. The Magic Pointer will be launched soon on Googlebook laptops. Experimental demonstrations also include capabilities such as pointing to edit images or finding locations on maps within Google AI Studio.
Anthropic restricts Pro/Max subscribers from using Claude models on open agent platforms. A Hugging Face blog post proposes two alternatives: one is calling hosted open-source models (such as GLM-5) via Hugging Face Inference Providers.
Google DeepMind has released Gemini 3.1 Flash Live, which is currently the highest-quality audio and voice model. It offers a Gemini Live API preview for developers, integration into Gemini Enterprise for Customer Experience for businesses, and public availability via Search Live and Gemini Live.
Cohere has released an open-source Automatic Speech Recognition (ASR) model named Transcribe (cohere-transcribe-03-2026), which adopts a Conformer encoder-Transformer decoder architecture.
Google DeepMind released new research that constructs and publicly shares the first empirically verifiable toolkit for measuring harmful AI manipulation, used to assess AI's ability to distort human beliefs and behaviors in real-world scenarios by exploiting emotional and cognitive vulnerabilities.