Moonshot AI just dropped Kimi-K2.7-Code and open-sourced it. This is their latest coding model, and the benchmark jumps over K2.6 are not small. If you care about open-source coding tools, this one deserves your attention.
What Is Kimi-K2.7-Code?
It is a Mixture-of-Experts model with 1.1 trillion total parameters, but only 32 billion are activated per token. That is the trick: you get the knowledge capacity of a massive model at the inference cost of a much smaller one. It has a 256K token context window and supports multimodal input (text and images) through a built-in vision encoder called MoonViT.
The license is Modified MIT, not vanilla MIT or Apache 2.0. That likely means some restrictions around commercial redistribution or competitive use. Worth reading the fine print before building a product on top of it.
How Much Better Is It Than K2.6?
Moonshot claims significant gains across every benchmark they tested. The headline numbers from their announcement:
- +21.8% on Kimi Code Bench v2 (their internal coding benchmark)
- +11.0% on Program Bench
- +31.5% on MLS Bench Lite (multi-language software engineering)
- +9.3% on Kimi Claw 24/7 Bench (agentic coding tasks)
- +9.5% on MCP Atlas (tool use)
- +11.4% on MCP Mark Verified (tool use, verified)
On top of that, it uses about 30% fewer reasoning tokens than K2.6. That means faster inference and lower cost per query. Less overthinking, same or better results.
How Does It Stack Up Against the Big Names?
This is where it gets interesting. Moonshot published comparison benchmarks against GPT-5.5 and Claude Opus 4.8. K2.7-Code does not win across the board, but it is competitive in ways that matter for open source.
| Benchmark | K2.7-Code | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|---|
| Kimi Code Bench v2 | 62.0 | 69.0 | 67.4 |
| Program Bench | 53.6 | 69.1 | 63.8 |
| MLS Bench Lite | 35.1 | 35.5 | 42.8 |
| MCP Atlas | 76.0 | 79.4 | 81.3 |
| MCP Mark Verified | 81.1 | 92.9 | 76.4 |
The standout number there is MCP Mark Verified. K2.7-Code scores 81.1, which actually beats Claude Opus 4.8 at 76.4. That is a tool-use benchmark, and it means this open-source model is better at calling tools and working with external systems than one of the best closed-source models on the planet. GPT-5.5 still dominates at 92.9, but the gap is not as wide as you would expect from a model you can run yourself.
On raw coding benchmarks, K2.7-Code trails both GPT-5.5 and Claude Opus 4.8. That is expected. But it is close enough to be genuinely useful, and the fact that it is open source with only 32B activated parameters means you can run it on hardware that costs a fraction of what the big models require.
What Is Kimi Code?
Kimi Code is Moonshot AI’s answer to Claude Code and OpenAI Codex. It is a terminal and IDE-based coding agent that can edit files, run shell commands, explore codebases, search the web, and spawn subagents. You can install it with one command:
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bashIt runs on Mac, Windows, and Linux. The K2.7-Code model powers it behind the scenes. If you are already using Claude Code or Codex, Kimi Code is worth a look, especially now that the underlying model just got a significant upgrade.
Why This Matters
The open-source coding model space is moving fast. A year ago, open-source models were a curiosity for coding tasks. Now they are legitimately competitive with the best closed-source offerings on tool use and agentic benchmarks. K2.7-Code is not going to replace Claude Opus or GPT-5.5 for the hardest coding tasks, but it does not need to. It needs to be good enough for most workflows, cheap enough to run at scale, and open enough that you are not locked into someone else’s API pricing.
On all three counts, it delivers. The 30% reduction in reasoning tokens alone makes it more practical for real-world agentic workflows where you are burning tokens on every tool call. And the Modified MIT license, while not fully permissive, is a lot more open than “pay us per token forever.”
Where to Get It
- Kimi Code: kimi.com/code
- API Access: platform.moonshot.ai
- Model Weights: HuggingFace (moonshotai/Kimi-K2.7-Code)
- GitHub: github.com/moonshotai
A 6x High-Speed Mode was also teased in the announcement but no details were shared yet. I will update this post when that lands.
