K2.7 Code Highspeed is the fastest way to run Kimi's June 2026 coding model: about 180 tokens per second on typical coding tasks and up to 260 on short ones. It's still in Kimi Code, but the API version retires by the end of December 2026. Here's what it does, what it costs and what to move to.
Closed to new users since 31 October 2026. Existing users can keep using it, but Moonshot plans to retire all K2-series models by the end of December 2026.
Still listed as kimi-for-coding-highspeed on the Pro plan and above, as of November 2026.
Moonshot points API users to Kimi K3.1 Pro. For speed, use K3.1 Pro Highspeed.
On typical coding tasks, rising to 260 on short ones (Moonshot).
Up to six times standard K2.7 Code, by Moonshot's figure.
Text, image and video input, with thinking always on.
Input and output, double standard K2.7 Code.
| Feature | K2.7 Code | K2.7 Code Highspeed | K3.1 Pro Highspeed |
|---|---|---|---|
| API model ID | kimi-k2.7-code | kimi-k2.7-code-highspeed | kimi-k3.1-pro-highspeed |
| Kimi Code model ID | (replaced by K2.8 Preview) | kimi-for-coding-highspeed | k3.1-pro-highspeed |
| Output speed | Standard | ~180 tokens/s, up to 260 (Moonshot) | ~113–150 tokens/s (measured) |
| Context window | 256K tokens | 256K tokens | 1M tokens |
| Thinking | Always on | Always on, no effort setting | Always on: low, high or max |
| API input / output | $0.95 / $4.00 | $1.90 / $8.00 | $4.00 / $20.00 |
| Cached input | $0.19 | $0.38 | $0.20 |
| Kimi Code quota | – | About 3× regular models | About 2× K3.1 Pro |
| Kimi Code plan | – | Pro and up | Pro and up |
| API status | Closed to new users; retiring by end of Dec 2026 | Closed to new users; retiring by end of Dec 2026 | Current |
API prices per million tokens from Moonshot's pricing page. Kimi Code details from its model table, November 2026.
It's the fastest model Kimi offers. A typical 2,000-token answer streams in about 11 seconds at 180 tokens per second.
Output speed of Kimi's fast options.
Renames, small fixes, boilerplate and tests where you want the answer in seconds and the task isn't hard.
When you iterate quickly, a fast model that can fix its own mistakes may beat a slower, smarter one.
For big refactors or tricky bugs, K3.1 Pro (or its Highspeed version) is much stronger and has a 1M context.
The API version retires by the end of December 2026. Build new integrations on K3.1 Pro.
On the Pro plan or higher, type /model and pick kimi-for-coding-highspeed. Each task uses about 3× the quota of regular models.
Use kimi-k2.7-code-highspeed at https://api.moonshot.ai/v1 until the K2 series retires.
Connect with a Kimi Code key to https://api.kimi.ai/coding/v1 and the model kimi-for-coding-highspeed.
Full API setup: Kimi API guide. Speed and cost trade-offs: K3.1 Pro Highspeed.
Released on 12 June 2026, K2.7 Code is a thinking-only coding model built for long tasks. Moonshot reported better success on runs of 4,000+ tool calls and 12+ hours.
Moonshot says it uses about 30% fewer thinking tokens than K2.6 on average, which cuts cost and waiting on simple jobs.
The weights are on Hugging Face at moonshotai/Kimi-K2.7-Code under a Modified MIT licence, so it stays usable after the API retires if you host it yourself.
K2.7 Code was replaced behind Kimi Code's default ID by K2.8 Preview on 11 September 2026. See all versions on the Kimi models page.