月之暗面开源 Kimi K2

2025 年 7 月 11 日，月之暗面发布 Kimi K2 的 Base 与 Instruct 权重。此前“Kimi”首先是一个消费产品名：用户打开网页，把长文档交给远端服务。K2 把同一个品牌写进 Hugging Face 模型目录，开发者面对的不再是一个已经完成的聊天界面，而是 FP8 检查点、推理引擎、聊天模板和工具调用协议。

K2 是混合专家模型，总参数约一万亿，每个 token 激活约 320 亿。架构包含 384 个专家，每步选择 8 个路由专家并使用共享专家；上下文长度为 128K。320 亿描述的是一次计算中参与工作的规模，不是模型文件只有 320 亿参数。部署仍要容纳或分布那一万亿参数的权重，MoE 节省的是每步计算，不会让存储和通信成本消失。

技术材料称模型在 15.5 万亿 token 上预训练，并使用扩展到这一规模的 MuonClip 优化方法处理训练稳定性。Base 面向继续训练与研究；Instruct 则提供开箱即用的对话和 Agent 能力。首发 Instruct 被明确描述为“reflex-grade”，不进行后来推理模型那种长时间思考。这个限定区分了快速工具执行与更长推理链，也防止后续 Thinking 版本的能力被倒写进首日。

Agent 能力只有接入工具循环后才可见。仓库推荐 vLLM、SGLang、KTransformers 与 TensorRT-LLM，并提供 OpenAI、Anthropic 兼容 API；但工具调用还要求推理引擎支持 K2 自己的解析逻辑。模型生成一个函数名和参数只是第一步，客户端必须正确解析、执行工具、把结果送回，再处理下一轮。基准里的“Agent”到实际业务之间，隔着一整套由部署者负责的运行时。

月之暗面公布了编码、知识与工具使用等评测，并将 K2 定位在开放模型第一梯队。这些结果来自发布方及其评测协议，不能推导出每种软件仓库和工具链上都领先。更重要的外部检验，是模型在长任务中是否稳定遵守格式、是否会调用错误工具、失败后能否恢复，以及稀疏架构在不同集群上的吞吐是否抵消部署复杂度。

许可证采用修改版 MIT，允许使用、修改和商业分发；当商业产品超过一亿月活或月收入超过 2000 万美元时，界面需要显著展示 “Kimi K2”。这使大多数开发者拥有宽松入口，也把品牌署名条件保留给超大规模服务。K2 的发布因此不是简单地把聊天机器人“开源”：它把产品背后的能力拆成权重、引擎、协议和许可证，交给外部重新组装。下载结束时，真正的工作才刚开始。

Moonshot AI released Base and Instruct weights for Kimi K2 on 11 July 2025. Until then, Kimi had primarily been a consumer product name: users opened a web page and sent long documents to a remote service. K2 placed the same brand in Hugging Face model catalogs. Developers did not receive a finished chat interface. They received block-FP8 checkpoints, serving-engine instructions, chat templates, and a tool-calling protocol.

K2 was a mixture-of-experts model with about one trillion total parameters and 32 billion activated for each token. Its architecture contained 384 experts, selecting eight routed experts per step in addition to a shared expert, with a 128K context window. The 32-billion figure described the capacity involved in computation on one step. It did not mean the model file contained only 32 billion parameters. Deployment still had to store or distribute the trillion-parameter weights. MoE reduced per-token computation without erasing storage and communication costs.

The technical materials reported pretraining on 15.5 trillion tokens and described MuonClip, an optimization method adapted to keep training stable at that scale. Base was intended for further training and research. Instruct supplied a more immediate chat and agent experience. The first Instruct release was explicitly described as “reflex-grade,” without the long thinking used by later reasoning variants. That qualifier separated quick tool-oriented behavior from extended inference and prevented later Thinking capabilities from being written backward into launch day.

Agent capability became visible only after connection to a tool loop. The repository recommended vLLM, SGLang, KTransformers, and TensorRT-LLM, and provided OpenAI- and Anthropic-compatible APIs. Tool calls, however, required the serving engine to support K2’s native parsing logic. Generating a function name and arguments was only the first step. A client had to parse the output, execute the tool, return the result, and manage the next turn. Between “agentic” on a benchmark and an operational business process stood a runtime owned by the deployer.

Moonshot published evaluations for coding, knowledge, and tool use, positioning K2 among leading open models. Those remained publisher results under defined evaluation protocols; they did not establish leadership on every repository and toolchain. More consequential outside tests concerned whether the model maintained formats over long tasks, chose incorrect tools, recovered from failure, and delivered enough throughput on different clusters to justify the complexity of sparse deployment.

The modified MIT license allowed use, modification, and commercial distribution. If a commercial product exceeded 100 million monthly active users or $20 million in monthly revenue, its interface had to display “Kimi K2” prominently. Most developers received a permissive entry while very large services retained an attribution condition. K2 was therefore not simply a chatbot being “open-sourced.” It decomposed the product’s capability into weights, engines, protocols, and a license for outsiders to reassemble. When the download finished, the real work began.