Guide
OpenClaw + Kimi K3: Setup, 1M Context, and What to Expect
Kimi K3 is Moonshot AI’s new 2.8-trillion-parameter, native-vision model. Here is the reliable way to connect it to OpenClaw without confusing API access with the later open-weight release.
The short answer
Kimi K3 works with OpenClaw through Moonshot's official provider plugin and OpenAI-compatible API. Install the plugin, restart the gateway, set MOONSHOT_API_KEY, verify that your account can see kimi-k3, then select the provider-prefixed model moonshot/kimi-k3. OpenClaw requires the provider prefix; a bare kimi-k3 is not the correct model reference.
This is a launch-day guide, not a benchmark victory lap. Moonshot describes K3 as its most capable model, but also says it still trails Claude Fable 5 and GPT 5.6 Sol overall. Read the official Kimi K3 announcement for the architecture, evaluations, pricing, and release schedule.
Why Kimi K3 is interesting for an OpenClaw agent
- One-million-token context: useful for large repositories, long research trails, and tool-heavy sessions.
- Native vision: K3 can reason over screenshots and visual interfaces as part of the same model.
- Agent-focused training: Moonshot emphasizes long-horizon coding, terminal use, repository navigation, and visual coding.
- Open-weight path: K3 is the first announced open 3T-class model, although the actual weights follow the hosted launch.
Connect Kimi K3 to OpenClaw
- Install the official provider with
openclaw plugins install @openclaw/moonshot-provider. - Create a key in the Moonshot developer platform.
- Run the official Moonshot onboarding flow and enter the key when prompted. It stores the credential for the gateway without putting the key in the command itself.
- Run
openclaw models list --provider moonshotand confirmkimi-k3is returned for your account. - Switch intentionally to
moonshot/kimi-k3, start a fresh session, and test a small read-only tool task.
openclaw plugins install @openclaw/moonshot-provider
openclaw onboard --auth-choice moonshot-api-key
openclaw gateway restart
openclaw models list --provider moonshot
openclaw models set moonshot/kimi-k3
openclaw models statusOpenClaw's current provider documentation requires the external Moonshot plugin and uses MOONSHOT_API_KEY. Because K3 launched after many installed catalogs, model discovery is the deciding check. If your installed OpenClaw version does not list it, update OpenClaw before inventing custom keys. The official model-provider reference is the source of truth for the version you run.
Pricing and context planning
Moonshot lists K3 API pricing at $0.30 per million cache-hit input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens. A 1M context window is a capacity, not a target: repeatedly sending an entire repository can still cost more and run slower than retrieval plus focused context.
Launch-day limitations
- The hosted API is live, while Moonshot says full weights will be released by July 27, 2026.
- Moonshot recommends 64 or more accelerators for deployment; “open weight” does not mean laptop-friendly.
- Preserve the model's reasoning history during a session and avoid switching models mid-thread.
- K3 can be overly proactive. Keep destructive tools approval-gated and write narrow task instructions.
Kimi K3 or Kimi K2.6?
Choose K3 when native vision, extreme context, or its stronger long-horizon coding behavior matters. Keep K2.6 when you need a mature integration, wider third-party availability, or predictable existing prompts. Our OpenClaw + Kimi guide remains the better K2.6 reference.
Frequently asked questions
Can OpenClaw use Kimi K3?
Yes. Install the official @openclaw/moonshot-provider plugin, restart the gateway, add Moonshot's API key, and use moonshot/kimi-k3.
Are Kimi K3 weights available now?
Not yet on launch day. Moonshot announced Kimi K3 on July 17, 2026 and says full weights will arrive by July 27. The hosted API is available now.
Is Kimi K3 a good local model?
It is open-weight, but it is not a typical desktop model. Moonshot recommends a deployment with at least 64 accelerators, so most OpenClaw users should start with the hosted API.