All Moonshot AI (Kimi) models available through the API

Choose a current Kimi model by its documented role, identifier and supported parameters. Preserve a migration test when replacing older Moonshot or Kimi routes.

Model families at Moonshot

The current catalog lists kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6, with distinct coding and general-purpose positioning. Official documentation.

Start with the application’s required input and accepted output. A coding-focused candidate is worth testing when the task involves repository changes, but a name does not prove correctness. Use a fixture with explicit checks and inspect the patch rather than evaluating only the explanation around it.

For knowledge work, require evidence from the supplied material. Include a case where the requested fact is absent and another where passages conflict. Those cases reveal whether the application can communicate uncertainty and preserve source distinctions. Keep that evaluation independent of a model’s maximum documented capacity.

When evaluating a high-speed variant, measure the full time to an accepted result. Time to initial output, total generation time and time spent repairing an invalid result can move differently. A published speed-oriented positioning establishes what to investigate; it does not establish your application’s end-to-end performance.

All models

Copy the exact identifier from the selected record and inspect status before using it. Preserve the provider endpoint with that identifier. A model name found in a hosted catalog elsewhere may refer to a different serving route and account relationship, even when the underlying model family sounds familiar.

Cheapest / strongest / longest-context

The live lists rank recorded attributes. Cheapest by input price does not mean cheapest per accepted coding task; longest context does not mean best evidence retrieval. A strongest recommendation requires a defined task and comparative outcomes. Use the data to reduce the evaluation set, then make the final choice from your own acceptance record.

Keep output policy fixed while comparing cost. If one candidate produces a lengthy explanation and another produces only the requested patch, the apparent efficiency difference may be caused by the prompt contract. Decide what the product needs, apply the same requirement and inspect the resulting usage.

Deprecations and replacements

The catalog marks older kimi-k2.5 and moonshot-v1 routes retired and directs migration to current Kimi models. Official documentation.

Search the application configuration, examples and scheduled job definitions for the old identifier. Replacing it in a visible request file may leave a background process unchanged. Preserve a list of consumers and verify the new identifier from each launch path.

Treat migration as a behavior change. Run the smallest passing request, then add the application’s required tools, output format and media input separately. Keep the previous evaluation fixture and compare accepted results. If a field is no longer valid, remove or adapt it according to the new model reference rather than suppressing the error and hoping the default is equivalent.

A retired model page can remain useful as a migration record, but it should not look like an active recommendation. Keep status and replacement guidance prominent and prevent new examples from selecting the old route.

Naming and versions

Kimi’s model parameter reference describes model-specific parameter constraints; reasoning controls should be selected for the actual model family. Official documentation.

Do not assume that a suffix or a larger generation number implies every feature of an earlier model remains unchanged. Record the exact identifier and relevant parameter choices together. When a deployment uses environment configuration, keep the selected value in safe startup metadata so an unexpected result can be traced to the actual model.

Avoid a migration that changes the model, client library, tool schema and prompt simultaneously. Establish a passing baseline, change one layer and keep a rollback decision available. That sequence makes a failure diagnosable and lets the evaluation explain which change improved or worsened the application.

Connect the final selection to Kimi billing categories, available capacity and the Python request baseline.

Use the AI API cost calculator to turn the model and workload you are considering into an estimate.

Last verified 12 Sep 2026 · Source ↗

Frequently asked questions

Sources

• Kimi API quickstart ↗
• Kimi model list ↗
• Inference pricing ↗
• Recharge and limits ↗
• Error reference ↗
• Organization management ↗
• Account and billing ↗
• Context caching ↗
• Batch API ↗
• Model parameters ↗

Last verified 12 Sep 2026 · Source ↗