China Token Surge & Open-Source Model Leadership
Key Questions
What is Kimi K3 and its significance?
Kimi K3 is Moonshot's 2.8T-parameter open-weight model, the largest to date, with capabilities approaching Claude Fable 5 and GPT-5.6. It has sparked market reactions but analysts view the selloff as irrational.
Which Chinese models are advancing to trillion-parameter scale?
At WAIC, domestic models including Kimi K3 and MiniMax's next-generation over 2T parameters are targeting trillion-scale. Alibaba previewed Qwen3.8-Max, a 2.4T multimodal model claimed second only to Claude Fable 5.
What funding milestones were achieved by Chinese AI firms?
DeepSeek completed its first funding round at 320-350 billion RMB valuation with a second round under discussion at rumored $71 billion. Facewall Intelligence raised over 5 billion RMB in six months.
How is China's open-source AI ecosystem performing globally?
China's open-source large models have surpassed 10 billion cumulative downloads worldwide, with over 70,000 models hosted in national communities. Qwen and MiniMax are gaining overseas revenue share from companies like DoorDash.
What new model releases emphasize video and scientific capabilities?
VideoChat3-4B outperforms GPT-5 and Gemini 2.5 Flash on video grounding with fully open weights and training code. CAS released PanShi 2.0 scientific foundation model and Zhou Bowen introduced the 'Shusheng·DuanYan' platform.
What is the status of Thinking Machines Lab's Inkling model?
Mira Murati's lab released Inkling, a 975B MoE open-weight model positioned as the West's strongest but lagging Chinese counterparts with a 63% hallucination rate.
How are investment patterns shifting in AI?
AI investment is consolidating toward top players, aided by ChiNext's fifth-set listing standards. Word-token economics and A2A paradigms are emerging as new frameworks leveraging China's electricity and open-source advantages.
What 3D AI company achieved major funding?
Meshy raised nearly $400 million in Series B at over $1 billion post-money valuation, with ARR growing 12x and 12 million users, validating commercialization in the 3D generation space.
中国大模型持续领先：WAIC现场国产模型集体向万亿参数冲刺，Kimi K3（2.8T）最大开源模型发布，综合能力接近Claude Fable 5和GPT-5.6。MiniMax下一代超2T，行业形成千亿性价比+万亿旗舰分层。多模态与语言路线之争成为焦点。DeepSeek完成首轮融资，估值约3200-3500亿；第二轮融资洽谈中，估值传闻710亿美元。面壁智能半年融资超50亿，MiniCPM下载3800万次，发布2B文本和VLA模型。德国开源主权AI模型Soofi S发布。Mira Murati的Thinking Machines Lab发布Inkling（975B MoE开放权重），西方最强但落后中国，幻觉率63%。7款手机端侧大模型获备案，苹果集成阿里千问。印度企业加速转向中国AI大模型。Sacks警告美国可能失去AI领先优势。Databricks CEO确认GPU短缺。前吉尔吉斯总理指出AI竞争已从模型转向产业化。Five Eyes警告及蒸馏指控引发关注。AI投资逻辑转向选头部，科创板第五套标准为AI企业打通退出路径。Mozilla报告确认开源AI接近闭源，但收入失衡，中国开源战略意图明确，竞争焦点转向agentic harness。商汤大装置在WAIC展示异构推理性价比超NVIDIA H系列1.25倍，日均Token服务2.42万亿，启动银河计划，今年首发算力卫星。词元经济理论兴起，中国电价优势、开源模型优势、词元出海和A2A成为新经济范式。阶跃星辰发布Step AOS智能体底层架构。中科院发布磐石2.0科学基础大模型。WAIC主论坛聚焦物理智能、科学智能、算力系统战和智能体治理。周伯文在WAIC提出'科学元认知时刻'，发布'书生·端砚'平台。新华网WAIC观察指出三大转向。最新：VideoChat3-4B在视频时间定位上超越GPT-5和Gemini 2.5 Flash，开源权重和训练代码全公开。Alibaba预览Qwen3.8-Max（2.4T多模态），声称仅次于Claude Fable 5，但缺乏基准和开源细节。中国大模型出海热度飙升，千问、MiniMax等海外收入占比高，DoorDash、爱彼迎采用中国模型。工信部宣布中国AI开源大模型全球累计下载量突破100亿次，国家级开源社区托管超7万模型。周伯文演讲强调AI在科学发现上的瓶颈，端砚平台蛋白质进化提升77%。WAIC官方总结提炼具身、兴业、善治三大主线。Qwen3.8-Max预览版上线，正式版即将开源。最新：Kimi K3引发的市场恐慌被分析师定性为情绪化反应，效率提升是长期利好。Qwen3.8-Max正式发布开放权重，但缺乏新基准。AI 3D生成赛道Meshy完成近4亿美元B轮融资（投后估值超100亿），ARR增长12倍、1200万用户，商业化验证扎实。