Moonshot

Cookbook examples for cookbook/90_models/moonshot.

Run examples with:

.venvs/demo/bin/python cookbook/90_models/moonshot/<example>.py

Reasoning

Kimi models reason before answering and return their thinking as reasoning_content, which Agno parses automatically and feeds back into the conversation on later turns.

Two parameters control that reasoning. Which one applies depends on the model generation, and parameters that do not apply are ignored by the API:

reasoning_effort

Controls how much the model thinks before answering. See Use thinking effort.

MoonShot(id="kimi-k3", reasoning_effort="low")

Kimi K3 accepts "low", "high" and "max".

It defaults to "max" when the parameter is omitted. That is a strong default: K3 will happily spend a minute or more reasoning before answering a prompt that does not need it. Lowering it to "low" cuts that dramatically — several times faster on simple prompts, with correspondingly shallower thinking.

So pick deliberately rather than leaving it unset:

• Leave it unset (or "max") for genuinely hard reasoning — proofs, planning, multi-step analysis.
• Set "low" for chat, summarization, formatting, tool-calling loops, and anything else where the answer is not the bottleneck.

use_thinking

Toggles thinking on the Kimi K2.x line, which reasons by default. See Use the Kimi K2 thinking model.

MoonShot(id="kimi-k2.6", use_thinking=False)  # faster, no reasoning_content

Leave it as None (the default) to use whatever the model does on its own. Note that thinking cannot be turned off on every model — Kimi K3 always reasons.

Structured output

Kimi returns structured data two ways, both driven by output_schema:

Native structured output constrains the response to your schema, so prefer it:

agent = Agent(model=MoonShot(id="kimi-k3"), output_schema=MovieScript)

JSON mode only guarantees the output is valid JSON, not that it matches the schema — it infers the shape from your field descriptions. Use it as a fallback where the json_schema path is not accepted:

agent = Agent(model=MoonShot(id="kimi-k3"), output_schema=MovieScript, use_json_mode=True)

Either way, Kimi only emits JSON objects — never a top-level JSON array. Wrap lists in a field on your model rather than asking for an array at the root. See Use JSON mode.

Media

Kimi accepts each media type differently, and Agno adapts automatically — you just attach images, files, or videos to the run:

Images are sent inline, so there is no upload step. Files cannot be attached inline (Kimi rejects the file content part), so each is uploaded, its text is extracted, and that text is injected into the message. Videos are uploaded and referenced by a Moonshot storage URL. After an upload the Moonshot file id is stored on the media object itself, so add_history_to_context does not re-upload the same media on later turns. See Use the Kimi vision model.

Examples