Kimi K3: the complete developer guide

Summary

Kimi K3 is presented as Moonshot AI’s 2.8-trillion-parameter open-weights model for long-horizon coding, end-to-end knowledge work, and deep reasoning, served through Together AI’s OpenAI-compatible API. Its architecture combines Kimi Delta Attention (KDA), Attention Residuals, and Stable LatentMoE, which activates 16 of 896 experts per token—roughly 2%—and supports a 1M-token context. The implementation guide shows Python calls for configurable reasoning effort, reasoning-free instant mode, streaming, multi-image vision, strict JSON-schema output, and function tools, while noting automatic prefix caching. It also covers production-oriented details such as generous max_tokens for reasoning-heavy structured output, a 100 MB request-body limit for vision, tool-call history handling, and deployment from US infrastructure.

Takeaways

1. 1

   K3 activates 16 of 896 Stable LatentMoE experts per token, roughly 2% of the expert pool, while Kimi Delta Attention and Attention Residuals support scaling to a 1M-token context.

2. 2

   Together’s API exposes low, high, and max reasoning_effort levels, with max as the default; reasoning can also be disabled through reasoning={“enabled”: False}.

3. 3

   For structured output, the full thinking trace is generated before the first schema-constrained token, so an overly small max_tokens value can truncate the JSON rather than the reasoning.

FAQ