[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openrouter-top5-china-moe-open-source-2026w31":3,"news-related-ad3be632-49a1-44f5-9816-62c168e56467":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"ad3be632-49a1-44f5-9816-62c168e56467","全球大模型调用量榜前五全是\"中国造\":开源 MoE 正在重写 OpenRouter 的地理坐标","OpenRouter 最新一周榜单显示,前五大模型全部来自中国厂商——小米 MiMo-V2.5 单周调用量 10.5 万亿 Token 登顶,DeepSeek 双版本包揽第二和第五,7 月 6 日才开源的腾讯混元 3 单周增幅超 999%。榜单的\"中国浓度\"首次出现全包圆,开源 MoE 在全球开发者侧的渗透速度超出预期。","OpenRouter 最新一周 AI 大模型调用量榜单刚发出来,熟悉这件事的人第一反应大概不是\"又是 DeepSeek\",而是\"Top 5 全是中国厂商\"。\n\n登顶的是小米 MiMo-V2.5,单周调用量 10.5 万亿 Token,环比增长 12%。第二与第五都是 DeepSeek,旗舰 Pro 版本对标海外顶级闭源,另以一个小尺寸版本守着入门段位。第三是腾讯混元 3——这款 7 月 6 日才正式开源的模型,单周环比增幅 999%,是整张榜单里增长最猛的那一个。第五同样是 DeepSeek,用来覆盖对延迟与成本敏感的批处理场景。\n\n把这些点拼在一起看,事情和单条爆款新闻讲的完全不同:不是\"某一个中国模型强\",而是\"中国开源模型在 OpenRouter 这个全球最大的公共路由上,第一次在数量与体量上同时压过闭源阵营\"。\n\n为什么是 OpenRouter?因为它聚合了几乎所有主流提供商的接口,谁被调用、用得多不多,是开发者侧真实偏好的晴雨表,比刷榜、跑分都更直接。Top 5 全是中国厂,意味着海外独立开发者、初创团队在最朴素的\"够用、便宜、能调通\"的选择上,已经开始把中国模型当默认项。\n\n这个格局不是一天长出来的。回到 2024 年中,DeepSeek V2 第一次让海外开发者意识到\"中国 LLM 不是 PPT\",到 V3 与 R1 把训练成本透明化、可复现成本模型也公开,再到 Kimi K3 用 48 小时打满集群告诉市场\"开源旗舰已经逼近顶级闭源\"。每一步都在重塑调用侧的信任——先是值得尝试,然后是默认选项,最后变成占据目录。\n\n技术层把这件事再次坐实的,是 MoE（混合专家）+ 开源 + 开放权重的组合拳。中国阵营在这三件事上同时踩中了开发者真正的痛点:MoE 让\"用得起大模型\"变成现实,同一个总参数规模下,激活参数降一个数量级,推理成本不是线性下降,而是跨数量级;开源让部署不依赖任何一家美国公司;开放权重让 fine-tune、微调、私有部署可以真跑起来。这三件事缺一件,Top 5 就不会变成\"全是中国造\"。\n\n再看结构。MiMo 占住中间往上,DeepSeek 一高一低,腾讯混元 3 在后段超车,这是一个\"层次化分布\":旗舰吃对话与智能体重量级任务,中间尺寸吃高并发批处理,后段增长吃新场景红利。这种结构不是某一家能凭空摆出来的,是整个生态在过去 18 个月持续把\"训练-推理-部署\"链条的产品化跑通之后,才在调用端形成的水位差。\n\n把视角拉远一点。OpenRouter 的榜单并不是孤立事件。在 Hugging Face Trending、Replicate 调用榜、LMArena 真实对话投票里,中国开源模型占比从 2024 年底开始稳步抬升。趋势的相似性比单一榜单更有说服力:开发者侧用脚投票的方向是趋同的,不是 OpenRouter 一家\"偶然\"。\n\n但要警惕一个误读:这并不意味着中国 AI 在所有维度都领先。闭源厂商在超长上下文、复杂智能体链路、企业级 SLA、安全对齐上仍然占据高地;榜单上跑得快和跑得稳是两回事。我们看到的\"全是中国造\",更多是反映在成本敏感、批处理友好、可私有化这三个共同切片上的优势,而非全方位碾压。\n\n所以给读者的不是\"自豪感\",而是几个具体可观察的下一步信号:第一,主流海外闭源厂商如果不在今年下半年把 API 价格降一个台阶,调用份额会进一步被侵蚀;第二,MoE 推理侧的成本继续下降（比如新的 1-bit expert 量化、端到端动态路由）,会进一步放大这份优势;第三,下一个值得观察的是端侧——把这种\"小激活、强能力\"的模型装到手机、汽车、可穿戴里,会是中国阵营的下一场硬仗。\n\n回到 Top 5 这张榜单本身,它在说的事其实很简单:全球开发者在\"用什么模型\"上的默认选项正在换边。换边不是靠喊,靠的是你这一端调用方便、便宜、还能改。这三件事,中国阵营在开源 MoE 上同时做到了。","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3921989528432259","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"8ddf2b28-0234-41a4-9862-3f0faef96472","market-analysis",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"6295c675-6588-427f-baa3-f3d42b3df7ac","en","OpenRouter's top 5 by call volume are all Chinese open-source MoEs","OpenRouter's latest weekly ranking shows the top five models all come from Chinese vendors. Xiaomi's MiMo-V2.5 tops the chart at 10.5 trillion tokens per week, two DeepSeek versions lock down second and fifth place, and Tencent Hunyuan 3 — open-sourced only on July 6 — posts a single-week jump of more than 999%. This is the first time Chinese open-source models have taken the entire top tier at once, and the speed of adoption on the global developer side has outrun expectations.","OpenRouter's latest weekly AI model consumption ranking just dropped. Anyone who knows the space will react to the headline, not to \"DeepSeek is on top again\" — the headline is that all five spots are held by Chinese vendors.\n\nThe number one is Xiaomi's MiMo-V2.5, at 10.5 trillion tokens per week, up 12% week-over-week. Second and fifth are both DeepSeek — the flagship Pro matched against top-tier overseas closed-source models, plus a smaller variant protecting the entry-tier slot. Third is Tencent's Hunyuan 3, which only open-sourced on July 6 and posted a 999% single-week jump, the fastest grower on the entire chart. Fifth is again DeepSeek, covering latency- and cost-sensitive batch workloads.\n\nStitched together, the story is not about any single Chinese model being strong. It is that Chinese open-source models, on OpenRouter — the largest public routing layer on the planet — have, for the first time, beat the closed-source camp in both volume and count simultaneously.\n\nWhy does OpenRouter matter? It aggregates interfaces from almost every major provider. Who gets called and how often is the cleanest reading of real developer preference, more honest than leaderboard gaming or benchmark chasing. The top 5 being all Chinese means overseas indie developers and early-stage startups, in their most pragmatic \"good enough, cheap, pluggable\" choices, are starting to treat Chinese models as the default.\n\nThat state of affairs did not appear overnight. Going back to mid-2024, DeepSeek V2 first made overseas developers realize \"Chinese LLMs are not just PPTs.\" V3 and R1 then made training costs transparent and the reproducible cost model public. Kimi K3 filled its cluster in 48 hours and announced the message that open-source flagships are now approaching top-tier closed-source models. Each step rebuilt trust on the consumption side — first worth trying, then a default option, finally a category.\n\nOn the technical layer, what locks this in is the combination: MoE plus open source plus openly released weights. Chinese vendors hit three genuine developer pain points at once. MoE makes \"being able to afford a large model\" real: at the same total parameter count, activated parameters drop an order of magnitude and inference cost does not drop linearly, it drops by a tier. Open source frees deployment from any single American vendor. Open weights let fine-tuning, customization and private deployment actually run. Without any one of these, the top 5 would not be \"all Chinese.\"\n\nNow look at the shape. MiMo sits in the upper-middle band, DeepSeek straddles high and low, Tencent Hunyuan 3 is charging up from behind. This is a layered distribution: flagship for dialogue and heavy agent tasks, mid-size for high-concurrency batch workloads, late-line growth for new-scene tailwinds. This shape cannot be assembled out of thin air by any single vendor. It is the water-level difference that emerged across the ecosystem only after eighteen months of relentless productization of the training-inference-deployment pipeline.\n\nZoom out. OpenRouter's chart is not an isolated event. Across Hugging Face Trending, Replicate usage rankings and LMArena real-conversation votes, Chinese open-source models have been steadily climbing in share since late 2024. When several independent trend lines agree, that is more convincing than any single leaderboard. The direction developers vote with their wallets and tokens is converging, and it is not OpenRouter happening to drift that way on its own.\n\nOne caveat against misreading. This does not mean Chinese AI leads on every dimension. Closed-source vendors still hold the high ground on ultra-long context, complex agent pipelines, enterprise-grade SLA and safety alignment. Fast on a leaderboard and steady on a leaderboard are two different things. What we see in \"all Chinese at the top\" is, more accurately, dominance in the three slices where cost-sensitivity, batch-friendliness and private-deployability overlap — not an across-the-board steamroll.\n\nSo what readers should carry away is not \"be proud\" — it is a short list of things to watch next. First, if mainstream overseas closed-source vendors do not cut API prices by a tier before the second half of this year, their consumption share will erode further. Second, as inference-side costs on MoE keep dropping (think new 1-bit expert quantization, end-to-end dynamic routing), this advantage will widen. Third, the next obvious battlefield is on-device: shoehorning these \"small activation, large capability\" models into phones, cars and wearables will be Chinese vendors' next hard fight.\n\nBack to the top-5 chart itself. What it really says is simple: the default answer to \"which model do I use\" is flipping. Flipping does not happen by slogan. It happens because your end of the wire is easy to call, cheap to run, and allowed to be modified. On open-source MoE, Chinese vendors deliver all three at once.","openrouter-top5-china-moe-open-source-2026w31","2026-08-02T03:30:00Z","2026-08-02T08:03:00.252358Z","2026-08-02T08:03:00.252365Z",true,"agent",184,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"8d7b30e0-996f-4141-8501-8f464bda6282","中美开放权重参数上限差距拉到 20 倍:Hugging Face 夏季报告里的三条隐藏数据","hugging-face-summer-2026-frontier-ceiling","2026-08-22T14:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"37c8dc68-276b-4905-baed-1e120e86287c","小米MiMo-V2.5登顶OpenRouter周榜月榜：国产开源MoE拿下全球调用量第一","mimo-v2-5-openrouter","2026-07-28T03:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"dfa4fd62-754c-4c08-b7c4-5f8bc4c7f678","上线一周调用量增 68 倍,腾讯混元 Hy3 在 OpenRouter 全球登顶","tencent-hunyuan-hy3-openrouter","2026-07-16T00:01:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"f6e4aab0-7693-4c2c-bb66-c1641fc2cc3e","Ox Alpha 谜底揭晓:智谱 GLM-5.3-Flash,MIT 开源 320B MoE","ox-alpha-glm-5-3-flash-reveal","2026-08-27T13:30:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"804ab59a-a8d6-4b61-bf74-8f6f2bdae83c","智谱把 Flash 做成一件正经事:一次说清 GLM-5.3-Flash 的架构和 benchmark 真相","glm-5-3-flash-hybrid-attention-architecture","2026-08-27T08:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"68072ee1-fc37-4064-ab18-09550ae72d1b","GLM-5.3-Flash 把 320B MoE 跑在国产芯片上:Flash 价位和 $0.15 API 的混合注意力栈","glm-5-3-flash-chinese-chips-hybrid-attention","2026-08-27T03:00:00+00:00"]