大模型前沿速递

China AI continues to surge: DeepSeek completes first financing of $7.4B at $52-59B valuation, with founder investing $2B personally. Moonshot AI open-sources Kimi K2.7 Code; MIIT policy targeting 30+ high-value scenarios and 75% urban 1ms latency by 2028; Hong Kong Stock Connect adds four AI companies; first large model medical device certificate issued. DeepSeek V4 nearly matches Claude at 1/10 cost; Chinese models account for 61% of OpenRouter token consumption. Robot traffic surpasses human traffic; 智元机器人 #1 in humanoid shipments. State-led $280B data center investment. Rio de Janeiro municipal model found mixing Qwen and Nex weights, highlighting open-source governance issues. Chinese recommendation systems using Ascend NPU. Industrial LLM deployment paths discussed. GLM 5.2 benchmarked against GPT-5.5 and Opus 4.8, showing strong front-end coding but speed/stability gaps. WeChat opens AI agent ecosystem. MiMo Code from Xiaomi emerges. Rio model shell confirmed as Qwen/Nex mix. SiliconFlow raises 20B RMB; Lin Junyang founds new lab with 13.5B RMB valuation; domestic model price cuts exceed 90% across DeepSeek, Xiaomi, Tencent; Rio model controversy deepens. Z.AI (Zhipu) releases GLM-5.2 with 1M context window, MIT license, MoE architecture. Zhongke Wenge (中科闻歌) goes public as 'general decision large model first stock'. Hong Kong Science Park and SenseTime plan 4万P+ computing center by 2030. New: Qwable-3.6-27B fine-tuned model based on Qwen3.6, strong local deployment. Alibaba's Qwen-Robot suite enters embodied AI. Korea's strategic response to US-China AI competition adds geopolitical dimension. China's securities regulator expands STAR Market fifth-set listing standards to AI large model industry, opening IPO channels for unprofitable AI companies. Microsoft reportedly considering DeepSeek as cheaper alternative for Copilot, validating cost advantage. New: Shanghai Stock Exchange issues specific guidance for AI large model companies on STAR Market green channel, requiring at least one large model product online with scaled application. Morgan Stanley highlights China AI ecosystem as $trillion opportunity; Ant Group reports 35B RMB R&D investment; GLM 5.2 tops unofficial benchmark surpassing Opus 4.8; Token call volume grows exponentially; Open-source AI models eating frontier article analyzes value shift to inference/edge/silicon. AI super apps (Alipay, WeChat) signal application layer activation. Today: Hong Kong plans largest domestic AI computing center. Beijing startup 算苗科技 tapes out 3D TokenPU chip for inference. Morgan Stanley reiterates China AI opportunity. 硅基流动 B轮融资; 百度智能云金融AI report; Goldman Sachs bullish on MiniMax M3. AIEC 2026 highlights shift from training to inference ecosystem competition. Token economy perspective gains traction. New policy: 七部门发文引导平台企业加强大模型和智能体布局; 上交所发布AI大模型企业上市指引, 要求至少一个大模型产品上线并规模化应用; 科创板第五套标准扩围。DeepSeek融资细节确认: 梁文锋个人出资200亿。GLM-5.2实测击败闭源模型。Also: Sumi 7B diffusion model open-sourced; Headroom token cost reduction project. New today: LiblibAI reaches $2B valuation, showing China AI app layer maturing. DeepSeek financing details article provides deeper analysis.