Moonshot launches Kimi K2.7 Code, its open-source coding model

Moonshot AI released Kimi K2.7 Code on June 12, and open-sourced it. It's a coding-specialized member of the K2 line: a 1 trillion parameter Mixture-of-Experts model with 32 billion parameters active per token, a 256K context window, and a Modified MIT license. It's available today through the Kimi API and Kimi Code at $0.95 in / $4 out per million tokens, with the weights on HuggingFace. It's also live on OpenRouter, hosted by Moonshot AI and Novita at the same price.

What's new over K2.6

From Moonshot's launch notes:

• Coding and agent gains: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite, all relative to Kimi K2.6.
• Less overthinking: about 30% lower reasoning-token usage than K2.6, so the same answers cost fewer tokens.
• Long-horizon coding: better instruction following and higher end-to-end task success on multi-step coding work.
• High-speed serving: a 6x faster mode was teased at launch and has since shipped in limited beta as Kimi K2.7 Code HighSpeed (more below).

The launch benchmarks

Moonshot's reported numbers, against Kimi K2.6, GPT-5.5, and Claude Opus 4.8.

One result stands out: on MCP Mark Verified, a test of multi-step tool use, K2.7 Code posts 81.1 and edges past Claude Opus 4.8 at 76.4. GPT-5.5 still leads the coding rows. These are vendor numbers though, two of the six benchmarks are internal, and none are independently reproduced yet, so read them as a launch claim, not a verdict.

A first independent score: ErdosBench

The first independent signal arrived from ulam.ai's ErdosBench, a math-proof benchmark. On a 14-problem smoke test, Kimi K2.7 Code placed 2nd of seven models, behind only Claude Fable 5 and ahead of GPT-5.5 at its xhigh setting, Qwen 3.7 Max, its own predecessor Kimi K2.6, Claude Opus 4.8, and Grok 4.3. It accepted 5 solved or settled results with 13 of 14 coverage and no unsafe partials, and scored near the top on decisive moves and known-theorem use. ulam.ai's verdict: it loses to Fable 5 mainly on coverage and slightly lower audit quality.

This is one small smoke test on hard math, not a coding leaderboard, so treat it as a signal rather than a verdict. It is a striking one all the same: a coding-specialized model ranking above GPT-5.5 and Opus 4.8 on open math problems.

HighSpeed mode is now in beta

The high-speed mode teased at launch has shipped as Kimi K2.7 Code HighSpeed. Moonshot reports up to 6x faster serving: around 180 tokens per second on coding tasks with median-length inputs, and up to 260 tokens per second on shorter-context work. It is rolling out to Kimi Code Beta Program members, Kimi API developers, and Kimi Business users, with access capacity-limited for now. No invite is needed, anyone who joins the Kimi Code beta has a shot at access. It is a serving-speed upgrade on the same model, not a new checkpoint.

We will keep adding independent scores from coding leaderboards like CursorBench and FrontierCode as they publish. Pricing, context window, and host availability live on the Kimi K2.7 Code model page.

Sources

• Moonshot AI launch announcement for the reported numbers and model details
• ulam.ai ErdosBench results for the independent smoke-test ranking
• Kimi K2.7 Code HighSpeed announcement for the high-speed serving mode
• Kimi K2.7 Code model page for pricing, context window, and host availability