Kimi models: which version should you use?

Choose around the task, then test

• Long-document analysis: evaluate K3 when your material does not fit the 256K alternatives. Its larger window is capacity, not a guarantee that every detail will be retrieved correctly.
• Repository changes and coding workflows: evaluate K2.7 Code alongside K3 using the same bug, tests and tool permissions. Consider HighSpeed when waiting time matters enough to justify its different rate.
• Short tasks without a thinking phase: K2.6 provides a documented non-thinking option. K3 and K2.7 Code are unsuitable if disabling thinking is a hard requirement.
• Screenshots and video: check the selected model’s input format before choosing on name alone. Image understanding does not mean image generation.

Our selection method is to build a small set of representative tasks, define a pass condition for each, and record answer quality, elapsed time, token use and failures. Repeat difficult cases before choosing a default. A model that succeeds once has not established a reliable success rate.

For example, a coding trial can require a correct change, passing existing tests, and no unrelated edits. A document trial can require a correct answer with a supporting passage. These are suggested evaluation criteria; we have not published comparative test scores for these models.

Use the Kimi pricing guide to estimate costs and the Kimi API guide for connection details.

Context size is not an answer-length promise

A context window has to accommodate the request and generated tokens. Instructions, conversation history, tools and supplied material all use space. Leave room for the answer and, where applicable, thinking. A large document may be better handled through retrieval or focused excerpts than by resending everything on every turn.

K3 exposes low, high and max reasoning effort, with max as its public API default. Start an evaluation with an explicit effort and output budget so the comparison is reproducible. See the K3 parameter documentation.

Kimi Code uses a separate model menu

Kimi Code is a different product from the public Open Platform. Its current identifiers include k3, k3-256k, kimi-for-coding and kimi-for-coding-highspeed. The current Code documentation maps kimi-for-coding to K2.8 Preview. That does not make it a public Open Platform model ID.

Code context allowances also depend on the plan: for example, K3 can be limited to 256K on Plus, while higher eligible tiers provide 1M. Use the Kimi Code model reference for that product. Do not copy its key, endpoint or alias into the public API example.

What happened to K2.5 and Moonshot v1?

The Open Platform catalog marks K2.5 and the Moonshot v1 series as discontinued on 31 August 2026, and the older K2 series on 25 May 2026. These notices concern the provider’s public API service. They do not establish the status of downloaded weights or another hosting provider.

When following an older tutorial, check the model ID, parameters and access product together. Replacing the name alone can leave incompatible thinking or sampling settings. Verify the current catalog and retirement notices before migrating.