Kimi K3

Kimi K3 is Kimi's flagship model for agentic coding and knowledge work. Open Interpreter reimplements the provider-recommended Kimi Code harness in Rust, so K3 gets the request shape, tools, thinking history, and defaults it expects inside a Codex-like interface.

Start with a Kimi Code subscription

Install Open Interpreter, open your project, and start the terminal UI:

curl -fsSL https://www.openinterpreter.com/install | sh
cd my-project
i

On Windows, install with PowerShell, then run i in your project:

irm https://www.openinterpreter.com/install.ps1 | iex

On first launch, choose Kimi For Coding, complete the Kimi sign-in in your browser, then choose Kimi K3. In an existing session, open /model and make the same selections. Start a new session when switching to K3 so it begins with a fresh prompt cache and thinking history.

If you already have a compatible Kimi Code API key, you can start directly:

KIMI_API_KEY="..." interpreter \
  -c 'model_provider="kimi-for-coding"' \
  -m k3

Open Interpreter selects kimi-code automatically for Kimi providers. You do not need to install or run the external Kimi Code CLI.

Use a Moonshot Platform API key

Kimi K3 is also available as kimi-k3 through the Moonshot Platform API:

MOONSHOT_API_KEY="..." interpreter \
  -c 'model_provider="moonshotai"' \
  -m kimi-k3

For a single non-interactive task:

MOONSHOT_API_KEY="..." interpreter exec \
  -c 'model_provider="moonshotai"' \
  -m kimi-k3 \
  "Review this repository and fix the highest-impact bug."

The Kimi Code harness is inferred for this provider as well. You can confirm or change the active harness with /harness.

Context and reasoning

Kimi currently recommends a fresh session when selecting K3 because switching models invalidates the existing prompt cache. K3 uses maximum reasoning effort at launch. Kimi Code subscriptions provide a 256K context window on Moderato and up to 1M tokens on Allegretto and higher plans.

See Kimi's current model configuration for entitlement and reasoning details.

Pricing

Kimi lists these Kimi Code subscription prices as of July 16, 2026:

Kimi's direct API pricing for K3 at launch is $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens, and $15.00 per million output tokens. Prices and plan entitlements can change; check Kimi's current membership pricing and K3 launch post before purchasing.

Computer use

Kimi K3 can use Open Interpreter's bundled QA skill to operate and test interfaces. Ask it to test a web app and it can use agent-browser in a real browser. Ask it to test a native desktop app and it can use trycua for computer use.

For example:

Run this app, test the sign-in flow like a user, and fix anything that breaks.

The normal sandbox and approval settings still apply. Keep the request specific, and review actions that interact with accounts or external systems.

ACP and Codex SDK

Use Kimi K3 from any supported editor or client by configuring it to launch interpreter acp; see the ACP guide and the current ACP client directory.

Existing Codex SDK integrations need only point their binary override at Open Interpreter:

-const codex = new Codex();
+const codex = new Codex({ codexPathOverride: "interpreter" });

Set model_provider to kimi-for-coding with model k3, or to moonshotai with model kimi-k3. See the Codex SDK guide for complete TypeScript and Python examples.