Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
Empirical tests by the UK AI Safety Institute (AISI) show that GLM-5.2 and DeepSeek-V4-Pro have approached the performance of Claude Opus 4.6, released 4–7 months ago, on narrow-domain cybersecurity tasks.
On July 19, 2026, Kunlun Tech released the Matrix-Game 3.5 world model, focusing on overcoming the object permanence problem by proposing geometry-aligned Patch-level long-term memory and PRoPE+Warped RoPE spatial-aware encoding.
Doubao (ByteDance) released the Seed Audio 1.0 audio creation model, targeting complete sound scenes. It jointly models human voice, sound effects, and ambient sounds within a unified framework, supporting precise timeline control via a single prompt (100ms intervals), natural generation in multiple languages (20+ types), zero-shot voice cloning, and consistent extension of long audio up to approximately 2 minutes per session. The model is now available on the Volcano Ark Experience Center.
Doubao has released the Seed Audio 1.0 audio creation model, targeting complete sound scenes. Under a unified framework, it jointly models human voice, sound effects, and environmental sounds. It supports precise orchestration of emotional dialogue and sound effects with a single prompt, time precision control at the 100ms level, zero-shot timbre extension (approximately 2 minutes per session), multi-emotion/multi-role performance using the same timbre, and natural pronunciation generation in over 20 languages. It is now available on the Volcano Ark Experience Center.
ByteDance has released the Seed Audio 1.0 audio creation model, designed for complete sound scenes. It jointly models human voice, sound effects, and ambient sounds within a unified framework, supporting precise orchestration of multiple elements via a single prompt, 100ms temporal control precision, zero-shot timbre generation and extension (approximately 2 minutes per session), multi-emotion/multi-role performance with the same timbre, and natural generation in over 20 languages. The model is now available on the Volcano Ark Experience Center.
The GitHub engineering team points out that while writing code was previously the most expensive part of small feature requests, discussing whether to write it is now the most costly. When AI can produce an initial patch within half an hour, "try first, then judge" is more efficient than repeated debate—the patch is not a deliverable but a probe used to verify the scope of changes, testing difficulty, and ownership costs; what is truly expensive is not generating code, but whether humans can quickly understand, review, and maintain it long-term.
Google DeepMind 发布 Gemini 3.5 Flash Cyber，一款基于 Gemini 3.5 Flash 微调的轻量级网络安全模型，专用于快速发现。
Marketplace Pulse cites an empirical study by Autopilotbrand.com on Alexa for Shopping: Among 1,963 non-brand queries, 63.9% of AI-recommended products did not enter the top 10 organic search results for the corresponding search terms, and 40.9% did not even appear on visible search result pages; only 14.3% were sponsored products on the current search page, and 83% of those already had organic rankings.
Dharma AI published an article on the Hugging Face blog, based on its arXiv paper (2604.14314), empirically showing that DharmaOCR scored 0.925 on the Brazilian Portuguese OCR benchmark.
Moonshot AI released Kimi K3, an open-source large language model with 2.8 trillion parameters, utilizing the Kimi Delta Attention and Attention Residuals architectures.
Hugging Face disclosed a production infrastructure intrusion fully driven by an autonomous AI agent system. The attacker exploited remote code execution and template injection vulnerabilities in dataset processing to laterally penetrate multiple internal clusters. Hugging Face used its proprietary LLM analysis agent to reconstruct and assess over 17,000 attack logs. It found that commercial API models blocked the analysis due to safety guardrails, ultimately switching to the open-source model zai-org/GLM-5.2 to complete all forensics locally.
The Moonshot AI team released PerceptionBench, a benchmark for 10 categories of atomic visual perception capabilities derived from real model failure cases. It includes 3,000 validation questions that require only visual judgment, without reasoning or external knowledge.
The Moonshot AI team released PerceptionBench, a multimodal large language model evaluation benchmark focused on atomic-level visual perception capabilities. It contains 3,000 validated questions covering 10 categories: visual relationships, counting, attributes, depth and 3D, localization, comparison, fine-grained recognition, context integration, OCR, and hallucination. All questions require only image-based answering, without relying on reasoning or external knowledge. The benchmark has been open-sourced, with code and datasets published on GitHub.
Moonshot AI has released Kimi K3, an open-source large language model with 2.8T parameters, native visual capabilities, and a context window of up to 1 million tokens. It employs the Kimi Delta Attention and Attention Residuals architectures, integrates Stable LatentMoE (activating 16/896 experts), and will release full model weights on July 27, 2026.
In a Hugging Face blog post, the IBM research team pointed out that treating model routing simply as a classification task has fundamental flaws. This conclusion is based on empirical data from the CodeAct agent on the AppWorld Test Challenge.
AI21 proposes an executor-orchestrator coding agent architecture, where the open-source MiniMax-M3 model executes numerous rollouts in parallel to explore codebases, and GPT-5.2 extracts key context.
Hugging Face has released Real World VoiceEQ, a new benchmark for evaluating the human quality of speech AI. It is constructed based on over 1 million human ratings across diverse demographics, scenarios, and acoustic environments, covering four major tasks: ASR, TTS, S2S, and speech understanding. The benchmark includes more than 15 evaluation dimensions and over 60 metrics. The current version contains 785,000 TTS ratings and 48,000 STS ratings, with all evaluations conducted through its speech-native platform, Kairos.
Thinking Machines released the open-source multimodal large language model Inkling, with approximately 1T parameters and a context window of up to 1M, natively supporting image, text, and audio inputs; simultaneously launching the lightweight version Inkling-Small (276B total parameters, 12B active parameters).
AMD officially released ROCm 7.14, with the core feature being the transition of the TheRock automated open-source build and release system into production, replacing previous preview versions from 7.9 to 7.13; it adds support for the AMD Instinct MI350 PCIe GPU.
Christian Crowley, a partner at a16z Crypto, wrote that traditional financial institutions (TradFi) do not seek to merge with DeFi, but selectively adopt parts of blockchain technology that optimize existing business operations—such as atomic settlement, programmable money, and on-chain reconciliation—to reduce costs, minimize operational risks, and strengthen customer relationships; while actively discarding native DeFi features like open access, pseudonymity, and trustless execution.
Google DeepMind has officially launched the live pilot of ATL Saathi, a Gemini-based web application that provides 24/7 lesson planning and training support to teachers at India's Atal Tinkering Labs.
Sakana AI, in collaboration with IT University of Copenhagen and Autodesk, published a paper in Nature Communications proposing the Smart Cellular Bricks system.
vLLM v0.25.0 has been officially released. Model Runner V2 becomes the default execution path for all dense models, and PagedAttention has been completely removed; a new Streaming Parser Engine is added to uniformly support tool calling and reasoning parsing.
Ollama v0.32.0 has been released. Running the `ollama` command now launches an agent by default, supporting chat, coding, web search, and task delegation, with `glm-5.2:cloud` as the default model; the Codex App integration has been renamed to ChatGPT and can be launched via `ollama launch chatgpt`; launching legacy agent models (such as CodeLlama, Qwen2.5-coder, Llama 3.x, Mistral, StarCoder, DeepSeek-R1) will display a deprecation warning.