Kimi K3 发布并开源

2026 年 7 月 16 日，月之暗面在 WAIC 前夕发布 Kimi K3：总参数 2.8 万亿、每个 token 激活约 1040 亿的混合专家模型，原生支持视觉，上下文窗口一百万。这类数字在发布日总是密集而抽象，真正值得记住的是另一件事——十一天后的 7 月 27 日，权重开放下载。从那一刻起，"全球参数最大的开源模型"和"首个登顶 WebDev Arena 的开源模型"同时成立，且都有文件可查、有许可可读。

架构上，Kimi K3 选择了稀疏路线：896 个路由专家加 2 个共享专家，每 token 只唤醒约 16 个。稀疏不是偷懒，是把总参数堆到万亿级之后仍然控制单次计算的办法。自研的 Kimi Delta Attention 是线性注意力的一种工程实现，让百万 token 的上下文不必配以平方级增长的算力；Moon Clip 二阶优化器则把训练数据需求从 40T 降到 20T。这些细节值得单独读，因为它们决定了开源之后别人拿到的到底是什么：不只是排行榜上的一行，而是可以复现、可以改造的工程资产。

开源不等于免费。Kimi K3 采用修改版 MIT 许可证：常规商用没有障碍，但连续 12 个月 MaaS 收入超过 2000 万美元的云服务商需要另行签约。这是 2026 年开源模型常见的护栏——权重开放，而大规模卖算力的商业模式另谈。发布当天，阿里云、华为昇腾、Nebius、Baseten、Fireworks，以及 vLLM、SGLang 等推理框架完成适配，说明产业侧早已在等待这个节点，而不是临时反应。

它改变的不是某一项测试的排名，而是选择的可能性。在此之前，想要一个"接近前沿且能自己部署"的模型，开发者要在小一号的开源权重与闭源 API 之间妥协；Kimi K3 之后，最强的可自托管选项第一次与"世界最大"同义。对数据敏感的企业、研究机构与个人开发者，这意味着前沿能力不必以交出数据为前提。

当然，规模与榜单不自动等于好用。稀疏模型的专家路由有它自己的失败模式，百万上下文的价值取决于真实任务是否用得上，修改版许可证的商业门槛也需要每个团队自己核算。这些留白，恰恰是开源模型的常态：它把一部分解释权交给了使用者。历史并不常在同一天给出"最大"与"开放"两个词，2026 年 7 月的这一周算一次。

On July 16, 2026, Moonshot AI announced Kimi K3 ahead of WAIC: a Mixture-of-Experts model with 2.8 trillion total parameters, about 104 billion active per token, native vision, and a one-million-token context window. Numbers like these blur together on announcement day. The detail worth remembering came eleven days later, on July 27, when the weights were released. From that point, "largest open-weight model in the world" and "first open model to top WebDev Arena" held simultaneously—and both were checkable facts with files and a license behind them.

The architecture is a sparse one: 896 routed experts plus 2 shared experts, with only about 16 awakened per token. Sparsity is not a shortcut; it is how you push total parameters into the trillions while keeping per-token compute bounded. The in-house Kimi Delta Attention implements linear attention so a million-token context does not require quadratic cost, and the Moon Clip second-order optimizer halves training data demand from 40T to 20T tokens. These details matter because they define what the community actually received: not just a row on a leaderboard, but engineering assets that can be reproduced and modified.

Open weights do not mean free. Kimi K3 ships under a modified MIT license: ordinary commercial use is unblocked, but providers whose MaaS revenue exceeds $20 million over 12 consecutive months must sign an additional agreement. That is the common guardrail for 2026-era open models—weights open, reselling compute at scale is negotiated separately. Day-one support from Alibaba Cloud, Huawei Ascend, Nebius, Baseten, Fireworks, and inference stacks like vLLM and SGLang suggests the industry had been waiting for this node, not reacting to it.

What changed is not a rank on one test but the range of options. Before, a team wanting a near-frontier model it could deploy itself chose between smaller open weights and closed APIs. After Kimi K3, the strongest self-hostable option and "largest in the world" are the same thing. For data-sensitive enterprises, research groups, and individual developers, frontier capability no longer has to mean handing over data.

Scale and leaderboards do not automatically equal usability. Sparse routing has failure modes of its own; a million-token window is only valuable when real tasks use it; and the license threshold is a number every team must run through its own business model. Those open questions are exactly what open models are like: some of the interpretation is handed to the user. History rarely hands out "largest" and "open" in the same week. In July 2026, it did.