Kimi K2.7 Code just dropped — a programming-specialized model from Moonshot AI, the company whose valuation hit $30 billion yesterday.
Core metrics: 30% reduction in reasoning token usage, performance close to GPT-5.5 on certain tasks, roughly one-third the cost. The demo: recreating a playable Minecraft clone in under 5 minutes.
The token reduction isn't simple quantization — it's dynamic depth control during inference. When the model reaches sufficient confidence in a reasoning branch, it stops further exploration and moves on. This doesn't require retraining, so cost savings are immediate.
K2.7 Code also includes a CodeSandbox — generated code runs in a sandbox first, auto-corrects syntax errors, then outputs final code. First-pass success rate jumped from 61% (K2.5) to 84% (K2.7). A 6x speed mode launches next week. Total test cost: about $0.96.
At a moment when Claude Fable 5 is locked behind US export controls and GPT-5.5 remains a rumor, Kimi K2.7 Code makes one thing clear: the ceiling for programming AI isn't determined by any single country.
Read the full bilingual analysis on my blog.
This article first appeared on my tech blog
Top comments (0)