<MCP.Playground/>
online

Kimi K2.7 Code

Runs MCP tools

Moonshot’s coding model from June 2026, the last of the K2 line before K3. It is open-weight, reports a score on MCP-Atlas, and always thinks, which the MCP client has to handle.

Vendor
Moonshot AI
Released
12 June 2026
Context
262K tokens
Input
Text, image

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

MCP Evals

Can Kimi K2.7 Code actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Kimi K2.7 Code drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Kimi K2.7 Code as the driver.

Run evals →See a sample

What Kimi K2.7 Code is

Kimi K2.7 Code is a 1T-parameter mixture-of-experts model with 32B active per token: 8 of 384 experts. It has a 256K window, takes text and image input, and is released under a modified MIT licence. Moonshot built it for end-to-end programming over long contexts and says it uses about 30% fewer reasoning tokens than K2.6. Thinking is forced on, and so is preserved thinking: the model expects its earlier reasoning to be sent back on every turn.

What Moonshot AI says

  • MCP-Atlas: 76.0.
  • Kimi Code Bench v2: 62.0, up from 50.9 for K2.6. Program Bench: 53.6, up from 48.3.
  • About 30% fewer thinking tokens than K2.6.
  • The coding numbers are on Moonshot’s own benchmarks. At release there were no independent SWE-bench Verified results.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • Kimi Code, Moonshot’s coding agent, which connects to MCP servers
  • Moonshot’s OpenAI-compatible API, so MCP clients that accept a custom endpoint can point at it with a base-URL change
  • The open weights, served with vLLM, SGLang or KTransformers, behind any MCP client
  • MCP Playground’s Agent Studio: paste your server URL above and run K2.7 Code against it in the browser, with no Moonshot key

What to watch for

  • Thinking and preserved thinking are both forced on. The model card says reasoning_content has to be kept in the history across turns. A client that drops it can make later tool calls in the same loop worse, the same failure as on Kimi K3.
  • Unlike K3, OpenRouter lists parallel_tool_calls for K2.7 Code. If your server has tools that can run independently, this is the Kimi that can call them together.
  • MCP-Atlas measures real MCP tool use, so the 76.0 is closer to your workload than most coding scores. It is still Moonshot’s number on Moonshot’s harness.

Ask K2.7 Code to fetch three unrelated things from the complex-schema mock in one turn, then run the same prompt on Kimi K3. K2.7 Code can issue the calls in parallel; K3 on OpenRouter may run them one after another. Count the turns each one takes. All the mock servers →

Available here

ModelModel ID
Kimi K2.7 Codemoonshotai/kimi-k2.7-code

Go deeper

Frequently asked questions

Can I turn off thinking on Kimi K2.7 Code?
No. The model card says thinking and preserve_thinking are forced on. Your client has to pass reasoning_content back between turns.
Kimi K2.7 Code or Kimi K3?
K3 is the larger, newer general model with a 1M window. K2.7 Code is smaller, coding-focused, has a 256K window, and lists parallel tool calls on OpenRouter. Both are available here, so run the same prompt on each against your server.
Is Kimi K2.7 Code open source?
The weights are on Hugging Face under a modified MIT licence. At 1T parameters it still needs a multi-GPU setup to serve.
Does Kimi K2.7 Code support MCP?
Yes. Moonshot reports 76.0 on MCP-Atlas, a benchmark of real MCP tool use, and it works through any MCP client that turns tools into function calls and keeps reasoning_content in the history.

Sources

Kimi K2.7 Code for MCP: Specs & Tool Calling — Test Free | MCP Playground