- Real-time web search across 100+ websites
- Analyze up to 50 files (PDFs, Docs, PPTs, Images) with ease
- AI slides & websites maker
- State of the art coding capabilities
- Enhanced image understanding beyond basic text extraction
This is the 13th launch from Kimi AI - Now with K2.6. View more
Kimi K2.7 Code
Launched this week
Kimi K2.7 Code is Moonshot AI’s latest coding-focused agentic model, built for long-horizon software engineering, 256K context, multi-step tool use, multimodal inputs, and around 30% lower reasoning-token usage than K2.6. Available in Kimi Code, Kimi API, and as open weights/code.
Free Options
Launch Team
DeepSeek
Flowtica Scribe
Hi everyone!
Kimi K2.7 Code is open-weights and focuses on improving real-world long-horizon coding performance. Compared with K2.6, it shows clear gains in instruction following over long contexts and higher success rates on multi-step coding tasks.
It also reduces overthinking quite a bit, with 30% lower reasoning-token usage. The model runs with thinking mode on by default and has better support for vision + tool calling in agent workflows.
Kimi Code has already upgraded its default model to K2.7 Code, and a 6x faster high-speed version is coming!
The 30% drop in reasoning tokens alongside better multi-step task success is the interesting signal here. It suggests you're pruning unproductive reasoning chains rather than just thinking less. We've seen agent costs spiral on complex multi-turn tasks because of runaway chain-of-thought. How did you train the model to distinguish productive reasoning steps from redundant ones?
The open-weights + 256K context combination is what I'd test first, especially on a repo task where the model has to keep tool outputs, diffs, and failed test logs straight. Lower reasoning-token usage is useful, but the tradeoff I wonder about is recovery after the agent makes a bad edit. Do you have evals that measure whether K2.7 can backtrack from a failed test run without losing the original instruction?
Interesting model. The 30% lower reasoning-token count is notable. Does that also reduce latency proportionally for typical multi-step tasks?
Interesting launch. For coding-focused models, the thing I’d want to test is not just generation quality, but how well it handles long-running repo work: keeping context clean, explaining risky changes, and recovering after failed tests.
Local Panel
To be honest, I really like Kimi, but this time the benchmarks are a bit below my expectations; they only seem to be slightly better than 2.6. But I really appreciate the fact that you’re open-source and constantly striving to improve. Thanks, team.