Take a coding task from brief to working code

Kimi K2.7 Code is built for end-to-end coding work, not autocomplete. You hand it a goal, such as fixing a failing test or building a small feature, and it plans the steps, edits the files, runs tools, and keeps working across a long session. The part that changes daily use: it takes images and video as part of the brief. Paste a screenshot of a broken layout or an error dialog instead of retyping what you see, and it folds that into the task.

It is an update of Kimi K2.6 focused specifically on coding, and the jump is measurable: on Moonshot's realistic coding suite the score goes from 50.9 to 62.0, with similar gains on agentic tool-use tests. It also spends about 30 percent fewer thinking tokens than K2.6, which means faster and cheaper runs on the same work. To be clear about where it stands, Moonshot's own comparison table still puts GPT-5.5 and Claude Opus 4.8 ahead on most benchmarks. The story here is an open model with an MIT-style license closing that gap at a fraction of the price.

The quickest path is OpenRouter: open the model page, start a chat in the browser, and give it a small real task from your day. Pricing there is about 85 cents per million input tokens, so an evening of testing costs pocket change.

If you already use a coding agent tool, point it at this model instead: Moonshot's API at platform.moonshot.ai speaks both the OpenAI and Anthropic formats, so most tools accept it as a drop-in backend. The weights themselves are public and ship pre-quantized, but at a trillion total parameters this is server hardware territory, not a home GPU project.

There is no demo Space, so trying it means an OpenRouter or Moonshot account rather than a one-click page. The benchmark numbers are Moonshot's own runs, and the closed frontier models still lead most of the published boards, so treat this as the best value tier rather than the best absolute coder. And while the weights are open, very few readers will run them; in practice you are using a hosted model that happens to publish its weights.

• Point it at a failing test or bug report and let it patch code until things pass.
• Paste a screenshot of a broken page and have it track down the layout bug.
• Build a small internal tool end to end from a written brief.
• Hand off repetitive chores like refactors, migrations, and dependency bumps.
• Swap it in as a cheaper backend for the coding tool you already use.

Try the demo

View model page

Read this issue on neuralexpedition.com →