Why Kimi K2.7 Code is different
The only open-weights model in this lab — and it out-tools a frontier closed model, with vision, at a fraction of the price.
Modified MIT license — the only open model here. Opus, GPT, and Gemini are all closed.
81.1 on MCP Mark Verified vs Claude Opus 4.8’s 76.4 — the agentic-tool benchmark.
Native MoonViT encoder handles text, image, and video — rare for an open coding model.
1T MoE / 32B active, thinking always-on, built for Cline, Roo Code, and Claude Code.
262,144 tokens — absorb an entire backend and refactor across dozens of files in one run.
$0.95 in / $4.00 out per 1M (~6× cheaper output than Opus), 30% fewer reasoning tokens than K2.6.
How it stacks up
| Model | Open? | MCP Mark | Input /1M | Output /1M | Multimodal |
|---|---|---|---|---|---|
| Kimi K2.7 Code | Yes | 81.1 | $0.95 | $4.00 | text/img/video |
| Claude Opus 4.8 | No | 76.4 | $5.00 | $25.00 | text/img |
| Claude Sonnet 4.6 | No | — | $3.00 | $15.00 | text/img |
| GPT-4o | No | — | $2.50 | $10.00 | text/img |
| Gemini 3.5 Flash | No | — | $1.50 | $9.00 | text/img |
Closed-model prices from the lab's pricing table. MCP Mark shown where published. A live head-to-head (run Kimi alongside GPT/Claude/Gemini) is planned for a later update.
🔑 Moonshot API Key
Get your API key at platform.moonshot.ai → · Keys auto-clear after 24 hours. 🔒 Forwarded to Moonshot API via our proxy — never persisted on our servers.
🌙 Select Kimi Models
Select up to 4 models. Kimi runs thinking always-on and uses fixed sampling, so there is no temperature control.
📋 Test Notes
- Kimi K2.7 Code is a 1T MoE coding specialist (32B active per token) with a 256K context window
- Sampling is fixed — temperature and top_p are ignored, so there is no temperature control here
- Thinking is always on; final answers appear in the response, reasoning is retained across turns
- API endpoint:
https://api.moonshot.ai/v1(OpenAI-compatible format) - Errors here don't affect the main lab. Pricing: ~$0.95 in / $4.00 out per 1M tokens