Vercel’s AI Gateway now serves Kimi K2.7 Code, Moonshot AI’s open-weight coding model — a 1-trillion-parameter mixture-of-experts architecture with 32B active parameters and a 262K context window.
K2.7 Code is built for long-horizon programming tasks and always runs in thinking mode. The key specs:
- 1T MoE / 32B active params — large capacity, manageable inference cost
- 262,144-token context — fits entire codebases in a single pass
- Native multimodal input — accepts text and images (pass a screenshot alongside your prompt)
- Always-on reasoning — no toggle; the model thinks through every request
- Benchmark gains over K2.6 — +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite
To use it through Vercel’s AI SDK, set the model to moonshotai/kimi-k2.7-code. That’s it — same interface you’d use for any other gateway model, with Vercel’s custom reporting, budget controls for API keys, and zero data retention support all applying.
The model targets frontend development, DevOps, and performance optimization specifically. Moonshot AI open-sourced the weights on Hugging Face, so self-hosting via vLLM or SGLang is an option if you want to skip the gateway entirely. Cloudflare Workers AI added the same model within days — distribution is moving fast.
NOTE
K2.7 Code is a coding-specialist model, not a general-purpose LLM. It’s tuned for agentic tool use and extended reasoning over code. For chat or content generation, the base K2.7 is the better pick.
Available now on Vercel AI Gateway — no waitlist, no new SDK version required. If you’re already routing through the gateway, swap the model string and test it against your current coding model on a real task.