Generate code using Moonshot AI provider integration

Compares two Moonshot AI (Kimi) models on a short summarization task, both authenticated with a single MOONSHOT_API_KEY: kimi-k3, Moonshot's flagship "thinking" model that reasons before answering (with showThinking: false keeping that reasoning out of the graded output), against kimi-k2.6, the cheaper previous-generation thinking model with reasoning explicitly disabled via the K2.x-only passthrough: { thinking: { type: disabled } } control. Grading uses plain icontains/icontains-any assertions rather than embeddings-based checks, since Moonshot doesn't expose an embeddings endpoint.

Use this to compare Kimi's current flagship reasoning model against its cheaper prior-generation model on the same task, or as a starting point for evaluating any Moonshot-hosted model. Leave temperature and other sampling parameters unset - Kimi models pin these to fixed values themselves, so setting them yourself has no effect. Model IDs rotate over time, so if a configured model 404s, swap in a current ID from Moonshot's own model list rather than assuming the example is broken.

npx promptfoo@latest init --example provider-moonshot
cd provider-moonshot

Requires only a MOONSHOT_API_KEY environment variable, obtained from the Kimi/Moonshot platform's API-keys console. Run the eval with promptfoo eval and inspect the side-by-side comparison with promptfoo view.

Integrates Moonshot AI's Kimi models (kimi-k3, kimi-k2.6) as Promptfoo providers, authenticated via MOONSHOT_API_KEY, with no other external service required since the example avoids embeddings-based assertions entirely.

Teams evaluating Moonshot/Kimi models who want a working comparison between a current flagship thinking model and a cheaper prior-generation one, using simple keyword assertions rather than an embeddings-dependent grading setup.