Moonshot AI Open-Sources New Generation Programming-Specific Model Kimi K2.7 Code
Moonshot AI has open-sourced the Kimi K2.7 Code programming model. Compared to K2.6, it shows significant improvements in long-context programming, instruction following, and agent autonomous execution capabilities. Code benchmark tests improved by 11%-31.5%, token consumption decreased by 30%, and API pricing remains the same as K2.6. The model requires thinking mode to be enabled for optimal performance.
What is Kimi K2.7 Code?
Kimi K2.7 Code is an open-source model developed by Moonshot AI, focused on programming and equipped with agent capabilities. It is stronger in programming and agent performance, especially showing significant improvements in long-range programming tasks within real-world projects. These improvements lead to higher end-to-end task success rates in complex software engineering workflows. K2.7 Code also enhances inference efficiency, reducing thinking token usage by approximately 30% compared to K2.6.
Benchmark Performance
We evaluated Kimi K2.7 Code against K2.6 on a set of internal and external benchmarks, covering two dimensions: programming capability and agent task execution.
On programming benchmarks, K2.7 Code showed significant improvements over K2.6: Kimi Code Bench v2 +21.8% (62.0 vs 50.9), Program Bench +11.0% (53.6 vs 48.3), MLS Bench Lite +31.5% (35.1 vs 26.7).
Stronger programming capabilities also brought stronger agent performance. On Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified, which measure autonomous agent task execution capabilities, K2.7 Code improved by approximately 10% compared to K2.6.
- Programming:
- Agent:
Kimi Code Bench v2 is an internal benchmark developed by Moonshot AI, and Kimi Claw 24/7 Bench is an internal benchmark for agent evaluation. Kimi K2.7 Code and K2.6 were tested via Kimi Code CLI with thinking enabled (temperature 1.0, top-p 0.95, 262,144-token context). GPT-5.5 was evaluated in Codex (xhigh), and Opus 4.8 in Claude Code (xhigh). For exceptions and full methodology for each benchmark, see the Hugging Face model card.
Built for Long-Range Programming
Software engineering in real-world projects is rarely completed in one step. Refactoring codebases, implementing features across multiple files, or debugging during long agent sessions require models to reliably follow instructions within long contexts and push tasks to final completion.
Kimi K2.7 Code is optimized for these long-range scenarios. Compared to K2.6, it follows instructions more reliable within long contexts and achieves higher end-to-end task success rates, making it more suitable for complex software engineering workflows.
Inference Efficiency Optimization
Reasoning models often tend to "think too much," consuming thousands of tokens on problems that do not require complex deliberation. Kimi K2.7 Code significantly reduces this tendency: compared to K2.6, it reduces average thinking token usage by approximately 30%.
On Kimi Code Bench v2, Program Bench, and MLS Bench Lite, Kimi K2.7 Code scored higher than K2.6 while consuming fewer tokens per benchmark.
For developers, this efficiency compounds continuously in every task: interactive programming sessions respond faster, production environment API costs are lower, and agent workflows can accomplish more work within the same context budget.
Model Architecture
Kimi K2.7 Code is built on a Mixture of Experts (MoE) architecture, with 1T total parameters, activating 32B parameters per token, supporting a 256K context length, and employing a Multi-head Latent Attention (MLA) mechanism, while integrating a MoonViT visual encoder with 400M parameters.
Model weights are open-sourced and available on Hugging Face.
How to Choose Between Kimi K2.7 Code and K2.6
Kimi K2.7 Code is built specifically for programming tasks. For general scenarios such as writing, analysis, and conversation, we recommend the more balanced K2.6.
How to Use Kimi K2.7 Code
Access Points
You can use Kimi K2.7 Code through the following methods:
- Kimi Code (https://www.kimi.com/code/zh). Kimi K2.7 Code is now the default model, with thinking mode enabled by default. To start using it, follow the configuration instructions on the page.
- Kimi API on the Open Platform (https://platform.kimi.com). Developers can call Kimi K2.7 Code via the Kimi API and integrate it into their own programming workflows, agents, and development tools.
Thinking Mode Requirement
Kimi K2.7 Code does not support non-thinking mode. Whether used via Kimi API or Kimi Code, it will always have thinking enabled. In Kimi Code, if a request disables thinking mode, it will automatically be served by K2.6.
Kimi K2.7 Code Pricing
Kimi Code Packages
If you wish to experience Kimi K2.7 Code directly through Kimi Code (including terminal and IDE plugins), you can choose our Code packages. The following prices are the monthly average when paid annually:
Each package includes usage quotas that refresh weekly. Higher-tier packages provide larger weekly quotas and higher concurrency limits, suitable for more complex projects. For the latest package details, please visit the official membership page.
Kimi API Pricing
Kimi K2.7 Code is available via the Kimi API, billed by usage and per token:
The API supports automatic context caching, which can reduce input costs for reused contexts (cache hit ¥1.30 per million tokens, cache miss ¥6.50 per million tokens). Prices exclude applicable taxes. For the latest rates, please refer to the official pricing documentation.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.