Moonshot AI Releases Kimi K2.7-Code: 21.8% Improvement on Code Bench v2 over K2.6

Moonshot AI has open-sourced a new specialized model, Kimi K2.7-Code. The model is oriented toward programming tasks and agentic work — when an AI system sequentially executes multi-step tasks without human intervention. It is distributed under the Modified MIT license and available through the Kimi API and Kimi Code service.

Architecture and Model Parameters

Kimi K2.7-Code is built on top of Kimi K2.6, released several weeks ago. The key difference is its fine-tuning for coding tasks: writing code to specification, debugging, automatic refactoring, and iterative work in development environments without manual intervention. The context window is 256K tokens. In practice, this means the ability to hold dozens of code files simultaneously in memory, full conversation history with a developer, or extensive technical documentation. In agentic tasks — when the model reads a file, modifies it, runs a test, reads the error, and applies a fix — such context is critically important.

The second key parameter is a reduction in reasoning-token consumption by approximately 30% relative to K2.6. Reasoning tokens are the model's internal reasoning before each response. In agentic scenarios, where the model reasons before each of dozens of steps, total consumption grows quickly. A one-third reduction is real savings in production load.

Results on Benchmarks

Moonshot compared K2.7-Code with its predecessor on six test sets and recorded improvements across all of them. The headline number is +21.8% on Kimi Code Bench v2. This is the company's internal test set, developed specifically to evaluate agentic coding capabilities: tasks requiring multiple iterations, file system work, and code execution.

• Kimi Code Bench v2: +21.8% improvement over K2.6
• Improvement recorded across all six benchmarks
• Reasoning-token consumption: ~30% reduction
• Context window: 256K tokens
• License: Modified MIT (commercial use permitted)

Moonshot does not publish absolute values on external benchmarks — only comparison with K2.6. This makes independent assessment of the model's position relative to Claude 3.7 Sonnet, Gemini 2.5 Pro, or GPT-4.1 difficult. Independent tests from the community will appear in the coming days: open weights allow running the model locally and conducting comparison.

Open Access and Ecosystem

Kimi K2.7-Code is released under the Modified MIT license. The license is commercially friendly: the model can be integrated into products, fine-tuned on proprietary data, and deployed on enterprise infrastructure. Open weights enable fine-tuning to specific code standards or rare programming languages.

Access is organized through two channels: Kimi API — for developers and companies embedding the model into their own systems and CI/CD pipelines, and Kimi Code — Moonshot's ready-made coding assistant, the equivalent of GitHub Copilot for end users.

The release of K2.7-Code fits a sustained trend: Chinese AI labs systematically open powerful coding models. DeepSeek Coder V2, the Qwen-Coder series from Alibaba, and now Kimi K2.x represent real competition to closed Western systems — often with more open usage terms.

What This Means

For companies automating code work through AI, Kimi K2.7-Code offers a combination relevant for production use: a large context window, reduced reasoning costs, and an open license. The ability to fine-tune makes the model attractive for teams needing customization to corporate standards or specific technology stacks.