Kimi K2.7 Code Comes to Mixlayer

Mixlayer Team·June 18, 2026

Kimi K2.7 Code is now available on Mixlayer.

Moonshot released K2.7 Code last week as a coding-focused model built on Kimi K2.6. It's aimed at the work our users ask about most: code agents, repository assistants, tool-driven workflows and automation around software projects. It thinks before it answers, always, and it has a 256K-token context window, enough to hold a real slice of a codebase.

Model ID: moonshotai/kimi-k2.7-code
Context: 256K tokens
Price: $0.75 per million input tokens, $3.50 per million output tokens

More than loading a checkpoint

Adding a model like Kimi touches most layers of our stack. We had to prepare the tokenizer, handle Kimi's special tokens, keep its control tokens out of your output, preserve the tool-call IDs it generates, run its mixture-of-experts layers efficiently, and make it behave the same as every other model through our APIs.

We kept that base work separate from our speculative decoding work on purpose. The first goal was reliable serving: load the model, keep the output clean, keep tool calls correct, and make it usable on the same API surfaces as the rest of the platform.

Sending it a coding task is a normal Chat Completions call:

curl https://mixlayer.ai/v1/chat/completions \
  -H "Authorization: Bearer $MIXLAYER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshotai/kimi-k2.7-code",
    "messages": [{"role": "user", "content": "Find the race condition in this function and fix it with the smallest change: ..."}]
  }'

K2.7 Code always thinks, so leave room in max_tokens for the reasoning as well as the answer.

Tool calls in Kimi's own format

Tool use is a big part of why you'd pick a coding model, and Kimi writes tool calls its own way, with dedicated section markers rather than the JSON or XML-style tags Qwen uses. So the gateway stays protocol-compatible, and modeld owns a parser that understands Kimi's native format. That parser also supports tool call streaming, so agents see a Kimi tool call as it's being written.

What comes next

Decoder CUDA graphs, speculative decoding and prefix-cache compatibility are separate layers that we'll add on top. Keeping them out of the launch kept it focused, and it leaves a clear path to lower latency once the generic speculative decoding protocol and Kimi's draft state are ready.

If you're building code agents on open models, try K2.7 Code against whatever you use today and tell us where it does well and where it doesn't. That's how we decide what to tune next.

The Kimi K2.7 Code guide covers reasoning behavior and prompting tips. Grab a key in the console, and see Moonshot's model card for the upstream details.