Moonshot AI Releases Open-Source Coding Model That Thinks 30 Percent Faster
Moonshot AI launched Kimi K2.7-Code, a one-trillion-parameter open-source coding model that cuts reasoning token usage by 30 percent, though practitioners note all benchmarks come from the company's own test suites.
Key Takeaways
- Kimi K2.7-Code uses 30 percent fewer reasoning tokens than its predecessor while scoring 21.8 percent higher on coding benchmarks
- The one-trillion-parameter model uses a Mixture-of-Experts design with 384 specialized sub-networks and a 256,000-token context window
- Released under a modified MIT license with API pricing at 95 cents per million input tokens and four dollars per million output tokens
- All published benchmarks come from Moonshot's own test suites and independent third-party evaluations are still pending
Chinese AI startup Moonshot AI has released Kimi K2.7-Code, an open-source coding model that uses roughly 30 percent fewer reasoning tokens than its predecessor. The one-trillion-parameter model is built on a Mixture-of-Experts architecture, often called MoE, which means it activates only 32 billion parameters at a time from a pool of 384 specialized sub-networks called experts. This design keeps the model fast and affordable to run while still delivering strong results on complex coding tasks.
What Makes K2.7-Code Different
The biggest improvement is token efficiency. When an AI model reasons through a coding problem, it generates internal thinking tokens before producing its final answer. These tokens cost money and take time, so reducing them matters. K2.7-Code cuts that overhead by 30 percent compared to the earlier K2.6 version, which directly translates to lower costs for developers running the model at scale. Moonshot reports a 21.8 percent improvement on its Kimi Code Bench v2 and an 11 percent gain on Program Bench. The model also supports a 256,000-token context window, letting it process entire codebases in a single session. For developers new to AI-assisted coding, this means the model can understand large projects without losing track of earlier files or context.
Open Source but Questions Remain
Kimi K2.7-Code is available on Hugging Face under a modified MIT license, making it free for most commercial and personal uses. API pricing sits at 95 cents per million input tokens and four dollars per million output tokens. However, some practitioners have raised concerns that all published benchmarks so far come from Moonshot's own internal test suites rather than independent third-party evaluations. Until the model is tested on widely recognized public benchmarks, its true competitive standing against rivals from Anthropic, OpenAI, and Google remains an open question.
The release reflects a growing trend in the AI industry where companies compete not just on raw capability but on efficiency and cost. As more organizations adopt AI coding tools for everyday development work, models that deliver strong performance at lower token costs could reshape how engineering teams choose their tools and how much companies budget for AI-powered software development.
Stay Informed
Weekly AI marketing insights
Join 5,000+ marketers. Unsubscribe anytime.