China AI Daily: Meituan's 1.6T LongCat-2.5, Big AI's Policy Token, GPT-6.2 Leak, and AI Rework Data
Domestic-trained 1.6T models and machine-readable training opt-out signals move AI practice in two concrete directions: model stacks that do not depend on imported chips, and a standards-based route for rights holders to exclude content from training pipelines. Western teams should also weigh the GitClear rework data and the ECB's valuation warning when scoping AI-assisted development and capital plans.
Tuesday's China AI briefing tracks two lines: domestic compute and AI governance. Meituan released LongCat-2.5, a 1.6-trillion-parameter multimodal MoE model trained entirely on Chinese domestic accelerators, which the press is calling a milestone for self-developed AI infrastructure. Separately, OpenAI, Google, Microsoft, and Anthropic jointly backed a new IETF-proposed standard called policy tokens: a machine-readable signal that lets content owners declare their material out of bounds for model training.
The briefing also flags model and developer-tool moves: OpenAI's GPT-6.2 Astra appeared in internal gray testing with about 19.8% of requests routed to it, skipping 6.1; HuggingFace open-sourced Carbon-X, a 14B DNA-language hybrid built on Qwen3-14B; DeepSeek released Super-Agent, an open-source programming agent that watches developer chats and gives unsolicited code feedback; Microsoft Ignite introduced new AI coding and reasoning models plus a quantum chip roadmap.
Market signals include GitClear data putting AI-generated code at 42% of shipped code with rework up 39%, the European Central Bank calling AI valuations clearly overvalued, OpenAI's round reportedly closing at only about $11 billion, and China's physical AI index up over 25% since October 15 with Unitree-linked names up over 50%. Wuzhen's Internet Light Expo is showing about 1,000 AI achievements and will release Zhejiang's 2026 AI development report.
A Chinese platform training a trillion-plus model on domestic accelerators suggests the country's compute stack has moved beyond feasibility into scale, so future releases will depend less on imported chips.
Policy tokens as a proposed IETF standard is the first time major US labs have rallied around a cross-vendor opt-out signal, giving copyright litigation and compliance a concrete technical answer.
GPT-6.2's gray-test rollout and the skip of 6.1 show release cadence is now a controlled experiment: about 20% request traffic is the validation cohort, and OpenAI can roll back if user satisfaction drops.
GitClear's 42%-generated, plus-39%-rework picture is the real cost line: AI raises output volume but also injects more revision loops, so teams should instrument review time and rework ratios before declaring productivity wins.
The ECB's overvaluation warning alongside China's physical AI index surge and OpenAI's shrunken round points to diverging risk appetites across geographies and sectors.