“Kimi K2 parameters” can mean model size or API request settings—they are not the same thing. Moonshot’s Kimi K2 technical page describes K2 as a mixture-of-experts model with 1 trillion total parameters and 32 billion activated parameters. Separately, the Kimi API quickstart documents request parameters for the specific model Kimi K2.6. Don’t assume those K2.6 settings apply to every Kimi K2 variant or API provider. (moonshotai.github.io)
Model size, API settings, and context length
- Model parameters describe the model’s scale. The official Kimi K2 page reports 1T total and 32B activated parameters for Kimi K2. (
moonshotai.github.io) - API parameters are request fields that control a completion, such as
max_tokensandthinking. The settings below are documented for Kimi K2.6. - Context length is a separate limit on the tokens a request can accommodate. The K2.6 quickstart states a 256K-token context window; that is not the same as the
max_tokensdefault. The documentation does not establish that one context figure applies to every K2 variant or provider. (platform.kimi.ai)
The official Kimi article about Kimi K2 Thinking discusses its own evaluation setup, including a 256K context setting. Evaluation settings are not automatically production API defaults or limits. (moonshotai.github.io)
Kimi K2.6 API parameter defaults and restrictions
The K2.6 API quickstart marks these request fields as optional and recommends using defaults rather than manually setting them. Its documented behavior is: (platform.kimi.ai)
| Request field | Documented K2.6 behavior |
|---|---|
max_tokens |
Defaults to 32,768. This is the documented default, not the context-window size. |
thinking |
Defaults to {"type":"enabled"}. The documented choices are {"type":"enabled"} and {"type":"disabled"}. |
temperature |
Fixed at 1.0 with thinking enabled and 0.6 in non-thinking mode. Other values result in an error. |
top_p |
Fixed at 0.95. Other values result in an error. |
n |
Fixed at 1. Other values result in an error. |
presence_penalty |
Fixed at 0.0. Other values result in an error. |
frequency_penalty |
Fixed at 0.0. Other values result in an error. |
Practical takeaway: don’t copy sampling settings from another model’s configuration. For K2.6, leave these fields out unless your integration needs them and the documented value is appropriate. In particular, setting an unsupported temperature, top_p, n, or penalty value may cause an error.
Thinking mode and tool calls
To disable thinking for K2.6, include "thinking": {"type": "disabled"} in the request body. When thinking is enabled and the request uses tools, the documentation permits tool_choice values of "auto" or "none"; other values result in an error. For multi-step tool calls, Kimi recommends retaining the assistant’s reasoning_content in the tool-call context to help maintain reasoning continuity and tool-call quality. (platform.kimi.ai)
The quickstart’s example uses the OpenAI-compatible Moonshot endpoint, the model ID kimi-k2.6, and a Bearer API key. Here is the same basic request pattern with thinking disabled:
curl https://api.moonshot.ai/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-d '{
"model": "kimi-k2.6",
"messages": [{"role": "user", "content": "hello"}],
"thinking": {"type": "disabled"}
}'
This is an example from the K2.6 documentation, not an independently tested request. Use it only when your deployment uses the documented Moonshot API endpoint and model ID; confirm the corresponding endpoint, model ID, and limits if you use a different K2 variant or provider. (platform.kimi.ai)
Which context limit should you use?
Treat context limits as variant- and provider-specific. Kimi’s K2.6 quickstart states a 256K-token context window, while the official K2 Thinking article’s 256K figure appears in its evaluation settings. Those statements do not by themselves define one universal limit for all K2 APIs. Before deployment, confirm the exact model ID and current context and request limits in the documentation for your chosen API provider. (platform.kimi.ai)
Bottom line
If you mean model size, Kimi reports 1T total and 32B activated parameters for Kimi K2. If you mean API settings, the documented parameter table here is specifically for Kimi K2.6: use its defaults, and follow its fixed-value and tool-use restrictions. These figures and settings describe different things—and should not be generalized across K2 variants without version-matched documentation. (moonshotai.github.io)