Moonshot AI · 2026-06-12 · major
Kimi K2.7-Code — Moonshot's 1T MoE Coding Model Beats Claude Opus on MCPMark
Moonshot's coding-tuned 1T mixture-of-experts model with 32B active params and 256K context. Scores 81.1% on MCPMark Verified (vs Claude Opus 4.8's 76.4%) and uses 30% fewer reasoning tokens than K2.6. Modified MIT, $0.95/$4 per Mtok.
Moonshot's coding-specialized fork of Kimi K2.6, with faster reasoning and a higher MCP tool-use score than Claude Opus 4.8.
Key specs
| Parameters | 1T |
|---|---|
| Active params | 32B |
| Context window | 256K |
| Price (input) | $0.95/Mtok |
| Price (output) | $4.00/Mtok |
| Experts | 384 (8 + 1 shared per token) |
| Mcpmark verified | 81.1% |
| Kimi code bench v2 | 62.0 |
| Program bench | 53.6 |
| Cache hit price | $0.19/Mtok |
Benchmarks
Pricing
| Input | $0.95 / 1M tokens |
|---|---|
| Output | $4.00 / 1M tokens |
| Cache hit | $0.19 / 1M tokens |
What is it?
Kimi K2.7-Code is Moonshot AI's open-weights mixture-of-experts model built specifically for coding and agent work. It has 1 trillion total parameters but only activates 32 billion per token, drawn from 384 experts (8 picked per token plus 1 shared). Context window is 256K tokens and the license is a modified MIT.
How does it work?
K2.7-Code keeps the K2.6 backbone but is fine-tuned for coding and tool calling. Moonshot says it cuts reasoning-token usage about 30% by curbing overthinking. It also adds a 'preserve thinking' mode that carries reasoning state across multi-turn conversations and a 'forced thinking' option. Vision input runs through the 400M-parameter MoonViT encoder shared with the rest of the K2 family.
Why does it matter?
It is the first open-weights coding model to beat Claude Opus 4.8 on MCPMark Verified (81.1 vs 76.4), and it narrows the gap to GPT-5.5 on Moonshot's own coding benchmark from 18 to 7 points versus K2.6. Teams running coding agents through API providers or self-hosting frontier-class weights get a credible alternative to closed flagships under a modified MIT license.
Who is it for?
coding-agent builders, self-hosters of frontier-class LLMs
Try it
huggingface.co/moonshotai/Kimi-K2.7-Code