Getting started with Kimi K2.7 Code for agentic coding
Why Moonshot’s Kimi K2.7 Code is a front-runner for agentic coding, the honest trade-offs against Claude Opus 4.8 and Qwen 3.7 Plus, and how to wire it into Ava Supernova via BYOK.
Kimi K2.7 Code from Moonshot is one of the strongest models available for agentic coding — the "decompose, execute, verify, iterate" loop that a real coding agent runs dozens of times per task. If you have not tried it on a real codebase, this guide covers the setup in Ava Supernova, plus the honest trade-offs against Claude Opus 4.8 and Qwen 3.7 Plus.
What makes it strong
K2.7 Code is Moonshot’s newest flagship — a large mixture-of-experts model tuned heavily for multi-turn tool use, exactly the pattern an agentic loop hits over and over. It plans decomposition well, recovers gracefully from failed tool calls, and holds a high ceiling on long-horizon refactors. On agentic-coding metrics it trades blows with the frontier closed models rather than trailing them.
Honest trade-offs
- Latency: slower than Qwen 3.7 Plus on short requests — for quick, chatty interactions, use something faster.
- Judgement calls: closed frontier models like Opus 4.8 can still edge it on one-shot architecture reviews and deep long-context analysis.
- Cost: cheaper than Opus 4.8 per token at list price; Moonshot offers volume pricing for heavier use.
- Availability: BYOK only in Ava — bring your own Moonshot key. It is not a managed-platform model.
Wiring it into Ava
K2.7 Code is BYOK. You need a Moonshot API key (platform.moonshot.ai) and a moment of setup:
- Open Ava — the VS Code extension, desktop IDE, or CLI.
- Go to Settings → API keys.
- Paste your Moonshot key into the Kimi slot.
- In the model picker, select Kimi K2.7 Code.
That is the whole setup. Ava routes your agentic work through it and falls back to the next model in your BYOK priority list if Moonshot has an outage.
When to reach for it vs Opus 4.8
K2.7 Code tends to win on multi-step agentic work; Opus 4.8 tends to win on judgement calls, one-shot reviews, and long-context analysis. If your work is mostly build-verify-iterate, start with K2.7 Code. For a single heavy review or a design critique, lean Opus 4.8. Both are one dropdown away — try the same prompt on each for any non-trivial task and compare.
No key? You are still covered
The managed platform runs Qwen 3.7 Plus as its default coordinator, a genuinely capable agentic model you can go a long way on without ever touching BYOK. K2.7 Code is for when you want a specific frontier model and are happy to hold your own key.