Kimi K3 Is Now Available in Kodo CLI

A New Frontier Model, Live in Kodo Today

We're adding Kimi K3 from Moonshot AI to Kodo's model lineup, available right now across the CLI, VS Code, Cursor, and JetBrains extensions.

$ kodo --model kimi-k3 "refactor this component to use the new data layer"

K3 landed just days ago and immediately jumped to the top of several independent leaderboards. It's not a minor addition — it's one of the most capable models we've plugged into Kodo, and it's priced aggressively enough that it changes the calculus for a lot of everyday work.

What Makes K3 Different

Kimi K3 is a 2.8 trillion parameter mixture-of-experts model — the largest open-weight model released to date. Despite the size, it stays fast: only 16 of its 896 experts activate per query, so you get frontier-scale capability without frontier-scale latency.

A few specs worth calling out:

1M token context window — enough to hold large codebases in a single session
Always-on "thinking mode" — K3 reasons through problems by default, not just when you ask it to
Modified MIT license — the same open-weight license family as the K2 generation, permitting commercial use and redistribution
Full weights land on Hugging Face July 27; until then it's available via API — which is exactly how it's wired up in Kodo today

Where It Ranks

On independent benchmark aggregation, K3 currently sits fourth among all frontier models overall — trailing only Claude Fable 5 and GPT-5.6 Sol, and edging out Claude Opus 4.8.

Where it really stands out is frontend work. K3 took the #1 spot on the Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587) — and it placed first in six of the seven scored frontend domains.

$ kodo --model kimi-k3 "build a responsive pricing table component with dark mode support"

If your day-to-day is UI-heavy — component work, layout, styling, interactive states — K3 is worth making your default.

The Price Is the Story

K3 is priced at $3 per million input tokens and $15 per million output tokens, with cached input dropping to $0.30/M. That puts it in the same tier as Claude Sonnet-class pricing, while scoring closer to the frontier tier on several benchmarks. Early third-party comparisons have described it as matching flagship-model output at roughly a third of the cost, in exchange for slower generation — a trade that's an easy call for anything that isn't latency-sensitive.

When to Reach for K3

Frontend and component work — its Frontend Code Arena results speak for themselves
Large-context tasks — the 1M window comfortably holds multi-file context
Cost-conscious agentic loops — cheap enough to run in /loop or /goal workflows without watching your credit balance
Anything that benefits from visible reasoning — thinking mode is on by default, so you get the "why" alongside the diff

Route it automatically with a model override:

{
  "defaultModel": "claude",
  "modelOverrides": {
    "components/**": "kimi-k3",
    "app/**/*.tsx": "kimi-k3"
  }
}

Getting Started

No update required — if you're already on Kodo, K3 is available immediately:

$ kodo --model kimi-k3 "explain this component tree and suggest a simpler structure"

New to Kodo?

$ npm install -g @eldlabs/kodo-cli
$ kodo init
$ kodo --model kimi-k3 "walk me through this codebase"

What's Next

We'll keep evaluating K3 as its full weights land on the 27th and update our benchmarks page as more independent numbers come in. As always, model selection in Kodo comes down to the same principle: use the best tool for the job, switch with a flag, never get locked in.

Questions about K3 or requests for what we should add next? Find us on [Discord](#).