Back Original

Kimi K3-256k

This page covers the models Kimi Code provides and how to switch between them in each client.

Model Overview ​

Kimi Code currently offers two models—Kimi K3 and Kimi K2.7 Code—across four model IDs, selectable by model ID in clients or third-party tools. Model specs:

Recommended model launch

k3-256k is now available. Within 256k context, it delivers the same results. k3 (1M) consumes about twice as much quota as k3-256k. Ideal for everyday Q&A, code completion, routine feature development, and single-file or small-file edits — video input is not supported.

💓 Reminder

Switching from K3 (1M) to K3-256k: When switching from k3 (1M) to k3-256k, if the current session's context already exceeds 256k, some coding tools such as Kimi Code CLI and Claude Code will perform a compact on the tool side.

Switching recommendations:

Switching from K3-256k to K3 (1M): When switching from k3-256k to k3 (1M), if k3-256k is close to the 256k limit and you don't want compact to lose information, you can switch directly to 1M. The current version switching from 256k to 1M does not affect the cache.

Model IDk3k3-256kkimi-for-codingkimi-for-coding-highspeed
Model versionKimi K3Kimi K3Kimi K2.7 CodeK2.7 Code HighSpeed
DescriptionKimi's most capable flagship coding model: 2.8T parameters, 1M context windowThe 256K context version of Kimi K3, effectively reducing consumptionGood at code completion and routine development tasksThe high-speed version of K2.7 Code, with the same coding ability and ~5–6× faster output
SpeedRegularRegularRegularHighSpeed (6× speed, 3× quota usage)
Context windowUp to 1M (for higher-tier members)256k only256k256k
Reasoningreasoning_effort:low / high / maxreasoning_effort:low / high / max (default high)Thinking:ONThinking:ON
AvailabilityAvailable to Moderato and above; 1M context for Allegretto and aboveAvailable to all Moderato members and aboveAll membersAllegretto plan or above
Multimodal inputImage, videoImage onlyImage, videoImage, video

Need a higher membership plan?

Different membership plans unlock different models, context windows, and speeds. Upgrade your plan →

Why did usage go up after the new model launched?

After switching models, the context cache built earlier no longer hits on the new model, so that context has to be re-prefilled. Usage therefore looks higher right after switching. Recommended action:

  • Start a new session when using the new model: this gives better results and lower consumption.
Why do I still get a 401 with the correct model ID?

When the requested capability exceeds your plan's entitlements, the server returns 401. Three common cases:

  • No K3 access: your plan is below Moderato and can't call k3, k3-256k — upgrade to Moderato or above.
  • No 1M access: on a Moderato plan, k3 supports up to 256K context; up to 1M context is available on Allegretto and higher tiers. k3-256k has a fixed 256K context limit.
  • No HighSpeed access: some plans don't include HighSpeed — upgrade to Allegretto or a higher tier to call kimi-for-coding-highspeed.

For the full error text and how to handle it, see the Error Reference.

Why isn't HighSpeed noticeably faster?

Two common reasons:

  • Mistyped model ID: the HighSpeed ID must be kimi-for-coding-highspeed; a wrong value silently falls back to the standard kimi-for-coding — no error, no speedup.
  • Tools and scripts dominate: HighSpeed only speeds up model output. Tool calls (reading/writing files, running commands, etc.) and script execution are unaffected, so when they take up most of a turn the overall speedup feels small.
How to reduce the overhead of switching reasoning effort?

Switching reasoning effort invalidates the context cache you've built up, so context that would have hit the cache must be re-prefilled. To avoid triggering re-prefill too often:

  • Pick an effort that fits the task and keep it consistent within a session;
  • When you genuinely need a different effort, start a new session rather than switching back and forth in a long session.

How to Switch Models ​

Usage notes

  • Start a new session when switching model IDs: switching models invalidates the context cache you've built up. We recommend starting a new session to get the best experience and avoid extra token consumption.
  • Fill in the Model ID, not the model version name: when calling a model, use one of the Model IDs from the table above (k3, k3-256k, kimi-for-coding, kimi-for-coding-highspeed). Entering a model version name like Kimi K3 or K2.7 Code will cause the call to fail.
  • K3 / K2.7 without Thinking routes to K2.6: keep Thinking on to use K3 or K2.7 Code; disabling thinking routes the request to K2.6.

Ways to switch to the target model:

Official clients ​

  • Official Kimi Code CLI: type /model to switch models—no config changes needed; if the latest model isn't listed yet, /logout and sign in again with /login.
  • Kimi Code for VS Code: pick the target model from the dropdown menu in the input bar; if it isn't listed yet, restart VS Code or reinstall the extension.

Third-party tools ​

Set the tool's Model ID to the target model. Detailed steps:

  1. Create an API Key in the Kimi Code Console.
  2. Fill in the Base URL and the corresponding Model ID in your tool.

Kimi Code API supports both OpenAI and Anthropic protocols. Base URLs:

ProtocolBase URL
OpenAI compatiblehttps://api.kimi.com/coding/v1
Anthropic compatiblehttps://api.kimi.com/coding/

For detailed setup steps, see the corresponding tool guide:

Before using K3 in third-party tools

K3's setup differs slightly from K2.7 Code. Before using it, check the two points below:

  • Context window: some tools default to a context window smaller than the max 1M—manually set the context-window field to 1048576 to use K3's full up-to-1M context.
  • Reasoning effort: K3 supports low / high / max; the effort a tool sends is mapped as below:
# default
null / undefined       → high
any other unknown       → HTTP 400 error

# → max
ultra / max / xhigh     → max

# → high (recommended)
high / medium           → high

# → low
low / minimum / light   → low

# → thinking disabled
none                    → thinking.type disabled