MoonshotAI: Kimi K2.7 Code

Overview. Kimi K2.7 Code is Moonshot AI's coding- and agentic-focused model, built on the Kimi K2.6 architecture and released as open weights under a Modified-MIT licence. It's a 1-trillion-parameter mixture-of-experts model with 384 experts and 32 billion active parameters per token, paired with a 262,144-token context window. Developers working on multi-step coding and agentic tool-use tasks that need a lot of context to reason over would reach for this model.

Key strengths. Its coding capability is the standout signal: Moonshot AI positions it explicitly as a coding-focused agentic model, and its measured coding benchmark score is markedly stronger than its general-quality score. It also carries a tool-calling loop suited to agentic workflows, and accepts image input alongside text. Extended thinking runs on every response, giving the model a mandatory reasoning pass before it answers — useful for harder coding and multi-step problems, though it comes at the cost of always being on.

When to use it. Use Kimi K2.7 Code for agentic coding tasks that benefit from a large context window — working across many files, long conversation histories, or extended tool-call chains. Its always-on extended thinking and large MoE footprint make it better suited to asynchronous or batch-style workloads than to latency-sensitive, real-time chat. It's a reasonable fit when the task is coding- or reasoning-heavy rather than general conversation, writing, or domain-specific work like legal or medical analysis, where it has no particular strength.

Limitations. Extended thinking cannot be turned off, so every request carries a reasoning trace and the added latency that comes with it. Moonshot AI has not disclosed a training data cutoff for this model, and its general-quality benchmark is modest relative to its coding strength, so it isn't a strong pick for broad, non-coding tasks. No content-safety or moderation tuning is documented for this model.

Sovereignty note. Kimi K2.7 Code is built by Moonshot AI, based in China; the creator does not carry CLOUD Act exposure. No Stav endpoint currently exists for this model, so endpoint-level data residency has not yet been established.