你可能没注意到，8月2日央视报道了一组数据，直接改写了全球AI格局表——OpenRouter最新一周全球大模型调用量榜单，前六名全是中国模型。登顶的，是小米。
不是OpenAI，不是Anthropic，不是Google。是一家做手机起家的公司，用一个叫MiMo-V2.5的模型，拿下了全球第一。单周调用量10.5万亿Token，是全球唯一一个单周突破10万亿的模型。第二名DeepSeek V4 Pro是6.37万亿，第三名腾讯混元Hy3是3.94万亿——小米一家的调用量，差不多等于第二加第三的总和。
凭什么是小米？
很多人看到"小米大模型全球第一"会本能地觉得不可思议。但拆解下来，逻辑很清晰。
第一是极致性价比。小米5月对MiMo系列API实施永久降价，最高降幅99%。输入价格低至1元/百万Token，缓存命中只要2分钱/百万Token。全球三五人的独立开发团队本来用不起Claude Opus，小米直接把价格地板砸穿了。
第二是精准定位Agent场景。MiMo-V2.5不是用来闲聊的，是专门给智能体场景设计的基座模型。万亿参数MoE架构，推理时只激活42B参数，成本极低但Agent能力极强。全球开发者用它写代码、跑自动化、执行长流程任务——这是"干活"的模型，不是"聊天"的模型。
第三是全系列MIT开源。所有模型采用MIT协议全量开源，支持自由商用和二次训练。独立开发者和初创公司可以直接下载部署，没有任何供应商锁定。
增长速度更夸张：MiMo-V2.5在4月23日公测时单周调用量才1.5万亿Token，两个月后暴涨到10.5万亿，增幅超过616%。雷军转发央视报道时只说了一句话："开源+极致性价比+Agent落地，全球开发者用脚投票。"
全球格局正在重写
完整榜单更值得细看：第一名小米MiMo-V2.5（10.5T），第二名DeepSeek V4 Pro（6.37T），第三名腾讯混元Hy3（3.94T，环比增长999%），第四名MiniMax M3，第五名DeepSeek V4 Flash，第六名智谱GLM-5.2。六款模型全部来自中国。中国模型全球市场份额达到63.5%，美国只剩35.5%。
这不是刷榜。OpenRouter汇聚了全球超过500万开发者，近一半用户来自美国，所有数据都是真金白银的付费调用。摩根大通预测，中国AI推理Token消耗量到2030年将再增长370倍。
当然需要加一个caveat：这份榜单只统计开源模型在第三方平台的调用量，企业私有化部署和美国大厂封闭API的流量不在其中。调用量领先不等于技术全面领先——闭源模型在综合能力天花板上仍有优势。但有一点已经无法否认：在全球AI的实用落地赛道上，中国已经跑出了领先身位。
当美国公司还在讨论AGI什么时候来的时候，中国模型已经深入到全球工厂、电商、编程、客服的真实生产线。小米这个全球第一，不只是一个模型的胜利，而是"开源+性价比+场景落地"这条中国路线的全球验证。
明天见。
You might have missed it, but data reported by CCTV on August 2 just rewrote the global AI scoreboard: on OpenRouter's latest weekly model usage chart, the top six spots all went to Chinese models. At the top: Xiaomi.
Not OpenAI, not Anthropic, not Google. A phone company, with a model called MiMo-V2.5, took #1 globally. Weekly usage hit 10.5 trillion tokens — the first model worldwide to break the 10T mark. #2 DeepSeek V4 Pro is at 6.37T, #3 Tencent Hunyuan Hy3 at 3.94T — Xiaomi alone accounts for roughly the sum of #2 and #3 combined.
Why Xiaomi?
"Xiaomi's AI model is #1 in the world" sounds improbable at first, but the logic is clear when you break it down.
First: extreme value pricing. Xiaomi permanently cut MiMo API prices in May by up to 99%. Input pricing dropped to 1 yuan/MTok, with cache hits as low as 0.02 yuan/MTok. Small indie dev teams globally that couldn't afford Claude Opus suddenly had the price floor smashed.
Second: Agent-focused positioning. MiMo-V2.5 isn't built for chit-chat — it's a base model designed specifically for agent workflows. Trillion-parameter MoE architecture activates only 42B parameters during inference, keeping costs minimal while delivering strong agent capabilities. Developers worldwide use it for coding, automation, and long-horizon task execution — it's a "get work done" model, not a "chat" model.
Third: full MIT open-source. All models use the MIT license, supporting free commercial use and retraining. Indie developers and startups can download and deploy with zero vendor lock-in.
The growth rate is staggering: MiMo-V2.5 was at 1.5T tokens/week when it launched April 23; two months later it hit 10.5T, a 616%+ increase. Reposting CCTV's report, Lei Jun said simply: "Open source + extreme value + Agent deployment. Global developers vote with their feet."
The Global Landscape Is Being Rewritten
The full chart tells the story: #1 Xiaomi MiMo-V2.5 (10.5T), #2 DeepSeek V4 Pro (6.37T), #3 Tencent Hunyuan Hy3 (3.94T, up 999% WoW), #4 MiniMax M3, #5 DeepSeek V4 Flash, #6 Zhipu GLM-5.2. All six from China. Chinese models hold 63.5% global market share; the US is down to 35.5%.
This isn't benchmark gaming. OpenRouter aggregates over 5 million developers globally, nearly half from the US, and all data represents paid API calls. JPMorgan forecasts China's AI inference token consumption will grow another 370x by 2030.
The necessary caveat: this chart counts open-source model usage on third-party platforms only. Private enterprise deployments and closed API traffic from US big tech aren't included. Usage leadership doesn't equal comprehensive technical leadership — closed-source models still hold advantages at the capability ceiling. But one fact is undeniable: on the practical deployment track of global AI, China has pulled ahead.
While US companies debate when AGI arrives, Chinese models are already embedded in real production lines across global factories, e-commerce, coding, and customer service. Xiaomi's global #1 isn't just one model's victory — it's global validation of the Chinese playbook: open source + value pricing + scenario deployment.
See you tomorrow.
当美国公司讨论AGI什么时候来的时候，中国模型已经跑在全球工厂、电商、客服的真实生产线上了。
—— Dawn Vision编辑部
While US companies debate when AGI arrives, Chinese models are already running on real production lines across global factories, e-commerce, and customer service.
— The Dawn Vision Editorial Desk
OpenRouter · 小米MiMo-V2.5 · 10.5万亿Token · 中国大模型 · DeepSeek · 腾讯混元 · 全球份额63.5% · 开源路线 · 性价比 · Agent落地
OpenRouter · Xiaomi MiMo-V2.5 · 10.5T tokens · Chinese LLMs · DeepSeek · Tencent Hunyuan · 63.5% global share · open-source playbook · value pricing · Agent deployment
Sources · 信源 Sources
- 央视8月2日报道
- 科技媒体 - OpenRouter周榜数据分析
本文基于 Dawn Vision 认知引擎处理的 9 个源信号生成，经编辑部人工审核。素材来源：央视新闻、科技媒体数据分析。
This article was generated by the Dawn Vision cognitive engine processing 9 source signals, with human editorial review. Sources: CCTV News, tech media data analysis.
