K2.8 Preview is the model behind Kimi Code's default kimi-for-coding ID. It has a 1M-token context on every paid plan, reads images and video, and is described by Moonshot as close to K3 with more efficient thinking. Here's what's confirmed, what isn't, and when to use it.
Rolled out to all Kimi Code users under the existing ID.
On every eligible plan, including Plus.
Paste screenshots, mock-ups or screen recordings.
“Close to K3” with more efficient thinking than K2.7 Code.
kimi-for-coding, so no settings change was needed.| Spec | K2.8 Preview |
|---|---|
| Released | 11 September 2026, in Kimi Code |
| Model ID | kimi-for-coding (the same ID K2.7 Code used before) |
| Status | Preview: no stable release date announced |
| Context window | Up to 1,048,576 tokens (1M) on every eligible plan |
| Input | Text, images and video |
| Thinking | low, high or max (default max); can be turned off |
| Speed | Regular speed |
| Plans | Kimi Plus and above (legacy: Andante and above) |
| Where | Kimi Code CLI, Desktop, VS Code and other tools using a Kimi Code key |
| Kimi API | Not listed on the pay-as-you-go Kimi API |
| Open weights | None published |
| Benchmarks | None published; Moonshot says performance is close to K3 |
From Kimi Code's model configuration page and What's New notes. Moonshot hasn't published parameter counts or a technical report for K2.8.
Kimi Code offers six model IDs. K2.8 Preview is the regular-speed everyday option and the only one with a full 1M context on the Plus plan.
| Model ID | Model | Context | Plans | Speed / quota |
|---|---|---|---|---|
k3.1-pro | K3.1 Pro | 1M (Pro plan and up) | Plus and up | Regular |
k3.1-pro-highspeed | K3.1 Pro Highspeed | 1M (Pro plan and up) | Pro and up | 2–3× speed, 2× quota |
k3 | K3 | 1M (Pro plan and up) | Plus and up | 1M uses 2× the quota of k3-256k |
k3-256k | K3 (256K) | 256K | Plus and up | Regular; images only |
kimi-for-coding | K2.8 Preview | 1M on all eligible plans | Plus and up | Regular |
kimi-for-coding-highspeed | K2.7 Code Highspeed | 256K | Pro and up | About 6× speed, 3× quota |
From Kimi Code's model table, November 2026. More on each: K3.1 Pro, K3.1 Pro Highspeed, K3 and K2.7 Code.
A simple way to split work between Kimi Code's models.
There are no formal benchmarks yet, only a few hands-on reports. Treat them as early signals, not proof.
| Date | Source | What they found |
|---|---|---|
| 11 Sep 2026 | Developer on X (@notjazii) | Same task as K3 at max thinking: K2.8 took 80 minutes and about $10 but needed more than one attempt; K3 got it in one go in 120 minutes for about $14. |
| 12 Sep 2026 | kimi-ai.free | Fixed a discount-calculation bug without changing the function signature, and the existing unit test passed. |
| Sep 2026 | Several reviews | Agree it's a sensible default for everyday coding, with K3 or K3.1 Pro kept for the hardest tasks. |
Single tasks from individual testers. Your results will depend on your codebase and prompts.
Install Kimi Code, sign in with a Plus plan or higher, and type /model. Choose kimi-for-coding, which is the default.
Point an OpenAI- or Anthropic-compatible coding tool at Kimi Code's endpoint with your Kimi Code key and the model kimi-for-coding.
Max is the default. Use high or low for simple tasks to get faster replies and save quota.
Full install guide: Kimi Code. Plan prices: Kimi pricing. In China, use api.kimi.com instead of api.kimi.ai.
Moonshot can update the model behind kimi-for-coding without changing the ID, so behaviour can shift. Re-test before relying on results.
Moonshot has only said performance is “close to K3”. There are no official scores and no independent leaderboard results yet.
There's no official pay-as-you-go API price or open weights. Third-party gateways that list it are unofficial.
With thinking turned off, Kimi Code serves K2.8 without thinking, even if you picked K3. A mistyped model ID also falls back to it quietly.