Moonshot AI released the open weights of Kimi K3.1 Pro (2.8T parameters, 1M-token context) on Hugging Face, while Tencent launched Hy4 with vision and computer-use abilities. Meta bought Conductor, a startup that runs multiple coding agents in parallel, and China issued a self-discipline convention for AI agents backed by Huawei, Alibaba, Tencent and Xiaomi. In Vietnam, Vietnamese-language Siri AI appeared in the iOS 27.3 beta.
New models
1. Moonshot releases Kimi K3.1 Pro open weights on Hugging Face
On Nov 4, Moonshot AI published the weights (MXFP4) of Kimi K3.1 Pro on Hugging Face: a 2.8T-parameter MoE with 104B active parameters, 1M-token context and multimodal input. By its own figures it scores 65.7 on Terminal-Bench 4.0, close to Claude Opus 5.5 (66.4), and Moonshot says it removed cyberattack material from training before the open release, making it the strongest current contender for open-model leadership.
2. Tencent launches Hy4, its first model with vision and computer use
Tencent Hunyuan released Hy4, its new-generation LLM succeeding the preview, with scaled-up training and, for the first time, image understanding and computer operation, plus better instruction following and long-text handling. It is available on Tencent Cloud Tokenhub, CodeBuddy, Yuanbao and via API on OpenRouter, showing the computer-use agent race has spread to Chinese labs.
3. Meta acquires Conductor, a startup that runs coding agents in parallel
Meta confirmed on Nov 4 it bought Conductor (terms undisclosed), an app that lets engineers run multiple Anthropic and OpenAI coding agents in parallel workspaces; CEO Charlie Holtz will work with Alexandr Wang and Nat Friedman at Meta Superintelligence Labs. Conductor's cloud service shuts down Nov 20, bolstering Meta's coding-agent push around Muse Code.
4. Revolut cuts AI coding costs by 90% by self-hosting GLM
Revolut switched its default from Claude Code to Z.ai's GLM-5.3 self-hosted on 56 B200 GPUs, cutting inference costs about 90% for roughly 2,000 engineers while keeping commercial models for vision tasks. It is a clear example of good-enough open models putting price pressure on commercial coding agents.
5. China issues a self-discipline convention for AI agent development
At the World Internet Conference in Wuzhen (Nov 4), 20 organizations led by the China Academy of Information and Communications Technology (CAICT), including Huawei, Alibaba, Tencent and Xiaomi, signed the 'Self-Discipline Convention on AI Agent Development' with eight groups of measures, from trust systems and interoperability standards to user protection. The firms pledged to put their agent development pipelines under oversight by industry associations and the public.
Xen advisory XSA-523 (Nov 3) credits Anthropic, using Claude, with finding CVE-2026-102703 and CVE-2026-102704, which let a malicious guest read and write other VMs' memory, including dom0, in all Xen versions from 4.6. Patches exist for newer branches; the case again shows LLMs becoming tools for finding serious infrastructure vulnerabilities.
7. OpenAI partners with Cathedral and sharply raises bug bounties
OpenAI announced a new security sprint, hiring Cathedral, a startup founded by DOGE alumni that previously accessed some of Anthropic's internal systems via its bug bounty, to attack OpenAI's systems using OpenAI's own models, and more than quadrupled rewards for critical bugs. The move follows Reuters reporting that OpenAI staff warned for years about internal security lapses.
8. Apple lists autonomous AI agents as a business risk
In its annual 10-K filing, Apple warned that autonomous agents could make unauthorized transactions, circumvent security or conceal actions, with consequences that may be hard to reverse, just as it opens Siri AI to act inside third-party apps.
9. GPT-6 Astra designs a hardware device in under a week
OpenAI said two employees used GPT-6 Astra to design an entire device for talking to the Codex agent, covering enclosure, PCB, firmware and UI, though several revisions were needed because the model missed manufacturing constraints. A newer internal model scores 0.875 on BenchCAD 2.0 versus Astra's 0.594, suggesting agents are moving into physical engineering.
10. Siri AI supports Vietnamese in the iOS 27.3 beta
Vietnamese became Siri AI's seventh language (built on Google's Gemini model) in the iOS 27.3 Beta, requiring an iPhone 15 Pro or newer; the beta targets developers and early testers. This follows Apple listing Vietnamese only as 'coming soon' the day before.
11. Over 539,000 accounts trial an AI platform for civil servants
At a Nov 4 press briefing, the Ministry of Science and Technology said its AI platform supporting public administration has been trialed across all 24 ministries and agencies and all 34 localities, with more than 539,000 accounts issued, showing LLMs entering government work at scale.