Dark Side of the Moon (Moonshot AI) launched quietly this weekKimi K2.7 Code——An open source model specially built for programming scenarios.Kimi The latest member of the K2 series, K2.7 Code achieved the highest coding benchmark results compared to its predecessor K2.6+21.8% improvement, while the reasoning efficiency is greatly optimized.
core competencies
K2.7 Code adoptionHybrid Expert Architecture (MoE), configure 384 experts, each token activates 8 experts plus 1 shared expert.
- Kimi Code Bench v2:50.9 → 62.0 (+21.8%)
- Program Bench:11% improvement
- Multilingual MLS Bench Lite: Increased by 31.5%
- Reasoning token consumption: Reduced by approximately 30%
Model passesKimi Code(Terminal/IDE Programming Agent) andMoonshot APIProvide services. The API is compatible with OpenAI and Anthropic SDK formats. You only need to modify the base URL to access.Modified MIT LicenseOpen source.
User experience/limitations
The biggest highlight of K2.7 Code isLong-range programming agent capabilities.$0.40/million input tokens, which is at a moderately low level in the programming-specific model. Considering that the open source weight can be deployed locally, the cost-effective advantage is obvious.
But the limitations are also clear: this is apure programming model, general conversation, creative writing, multi-language translation and other scenarios are not its strengths.Kimi Although Code Bench v2 significantly surpasses its predecessor, it is comparable to GPT-5.5 (69.0) andClaude Opus 4.8 (67.4) still has a gap.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 7.5 / 10 | Excellent programming scenarios, but limited general capabilities |
| 易用性 | 8.0 / 10 | Compatible with mainstream SDKs,Kimi Code built-in Agent experience |
| Cost-effectiveness | 8.5 / 10 | Open source + API dual track, reducing programming token consumption by 30% |
| 中文支持 | 9.0 / 10 | Developed by a Chinese team, the natural advantages of Chinese programming scenarios |
| 输出质量 | 8.0 / 10 | Programming benchmarks have improved significantly, but cutting-edge models still have gaps |
Overall rating: 8.2/10
Kimi K2.7 Code is a well-positioned open source programming model - it does not pursue omnipotence, but takes the long-range programming agent scenario to the extreme.Kimi Code, this upgrade is worth trying immediately.
Want the latest reviews of all AI models? AI Dash — Discover the best AI tools.