KimiFormatter

Formatter for Kimi (Moonshot AI) models (kimi-k3, kimi-k2.7-code, kimi-k2.6, kimi-k2.5, moonshot-v1 series).

Adapted to the latest Kimi open platform Chat Completions API (https://api.moonshot.cn/v1/chat/completions):

• On kimi-* models, temperature / top_p / n / frequency_penalty / presence_penalty are fixed by the platform and the API rejects requests carrying other values; these parameters are stripped from the request. The moonshot-v1 series still accepts them, so they are kept there.

• reasoning_effort ("low" / "high" / "max", default "max") is only supported by kimi-k3; it is stripped for other models

• thinking_budget is not a Kimi parameter and is always stripped; K2.x thinking is controlled through the thinking body parameter instead (see below)

• Only max_completion_tokens is documented; max_tokens is mapped to max_completion_tokens when the latter is not set. Reasoning tokens count towards max_completion_tokens, so the official guide recommends max_completion_tokens >= 16000 for thinking models

• tool_choice supports auto / none on all models; required is degraded to auto on the K2.x series; forcing a specific function is incompatible with thinking enabled (HTTP 400), so it is degraded to auto on always-thinking models (kimi-k3, kimi-k2.7-code) and on kimi-k2.6 / kimi-k2.5 unless thinking is explicitly disabled

• The strict parameter in tool definitions is not documented by Kimi and is not sent

• Assistant reasoning_content is preserved in the request history (inherited from the base message converter), as required by Preserved Thinking on kimi-k3 / kimi-k2.7-code and by thinking.keep = "all" on kimi-k2.6

Thinking mode: kimi-k2.6 / kimi-k2.5 accept a thinking body parameter (e.g. {"type": "disabled"}), which can be passed through GenerateOptions.additionalBodyParam("thinking", Map.of("type", "disabled")). kimi-k3 uses the top-level reasoning_effort option (GenerateOptions.reasoningEffort) instead. JSON mode is enabled via response_format = {"type": "json_object"}; json_schema is not supported.

Usage:

`OpenAIChatModel.builder()
.formatter(new KimiFormatter())
.modelName("kimi-k3")
.baseUrl("https://api.moonshot.cn/v1")
.apiKey(apiKey)
.build();
`