Kimi K3 with CC Switch and Codex CLI

Quick comparison

1. Quick browser demo with Puter.js

Puter’s Kimi K3 documentation shows how its browser SDK can run K3 without embedding your own API key in the page. Server-side use has a separate token-based authentication flow, so never expose a server token in browser code.

Load the browser SDK

<script src="https://js.puter.com/v2/"></script>

Call Kimi K3

puter.ai.chat("Explain this task in simple steps.", {
  model: "moonshotai/kimi-k3"
}).then((response) => {
  console.log(response.message.content);
});

Stream output

const stream = await puter.ai.chat(
  "Summarize this long task step by step.",
  { model: "moonshotai/kimi-k3", stream: true }
);

for await (const part of stream) {
  if (part?.text) process.stdout.write(part.text);
}

Use Puter.js for experiments and prototypes. Review current terms, quotas, supported modalities, and production requirements before building an automated workflow.

2. Official Kimi chat

1. Open kimi.com and sign in.
2. Use the model switch above the chat box and select K3.
3. Choose the available thinking strength: Low, High, or Max.
4. Start with normal chat tasks before testing large files or agent workflows.

Kimi’s help guide states that K3 and K3 Swarm consume credits. K3 Swarm is designed for large search and batch jobs, so check credits before using it. This is consumer chat access, not a general-purpose API key for Codex CLI.

3. Temporary ZenMux free API and CC Switch

ZenMux currently lists the following free-suffixed model. It is marked as free for a limited time, rate-limited, and PAYG-only:

moonshotai/kimi-k3-free

1. Create a ZenMux account and review billing, PAYG eligibility, and usage alerts.
2. Create a ZenMux API key, then keep it only in CC Switch or a secure operating-system credential store.
3. In CC Switch’s Codex tab, add a Custom OpenAI-compatible provider with the values below.

4. Enable the provider. If CC Switch uses Chat Completions for Codex, start Settings → Routing → Local Routing and enable the Codex integration.
5. Close existing Codex sessions, run codex again, and begin with a read-only prompt.
6. Before every important run, confirm the model still ends in -free. The paid model moonshotai/kimi-k3 is different and can create charges.

Keep a usage or spending alert enabled, even when the current model card shows free pricing.

4. Coding-focused alternative

For software-engineering workloads, check the separate Kimi model moonshotai/kimi-k2.7-code. It is not Kimi K3. Verify its current availability, price, and API requirements before use.

5. Future self-hosting

Moonshot has announced full Kimi K3 weights for 27 July 2026. Self-hosting can offer more control and privacy, but the full 2.8T-parameter model requires serious multi-GPU infrastructure. Do not plan to run it on a normal laptop, budget VPS, or typical gaming PC.

Paid Kimi K3 alternatives

Run a local Kimi-derived model with LM Studio

This is separate from the Kimi K3 API and Codex CLI. LM Studio runs downloaded model weights locally; it does not automatically connect a local model to CC Switch.

LM Studio for abliterated local models

LM Studio is a local-model app commonly used to load GGUF versions of community models, including variants labelled abliterated. In this context, abliterated usually means the uploader has modified or removed some refusal and safety behaviour in a model. That is why people often associate LM Studio with “jailbreak” models, even though LM Studio itself is only the local model runner.

1. Install LM Studio

Install LM Studio on Windows, macOS, or Linux. In Discover, search compatible models or paste a complete Hugging Face repository URL. LM Studio supports compatible GGUF downloads and imports.

2. Verify a model before downloading

Check the Hugging Face model card—not only its title. Confirm that actual GGUF files are present, the base model and uploader are identified, the licence is clear, the required RAM/VRAM is documented, and the repository has credible community feedback.

Some community uploads are labelled abliterated, meaning the uploader may have changed refusal behaviour. They are unofficial, untrusted derivatives. Do not use them to evade safeguards, and do not connect an untested model to terminals, private documents, business automations, or production systems.

3. Choose a realistic model size

These are rough starting points; use each model card’s requirements first. A Q4_K_M GGUF is often a practical quality-versus-memory starting point. Kimi K3 is a 2.8T mixture-of-experts model and is not a normal laptop or gaming-PC workload. Prefer a documented smaller Kimi-derived model on consumer hardware.

4. Load and evaluate safely

1. Download every required GGUF part and ensure enough storage is available.
2. Open LM Studio’s Chat, load the model, and start with a 4,096–8,192-token context window.
3. Let LM Studio choose GPU offload initially and close memory-heavy apps.
4. If loading fails, use a smaller quantization or model instead of forcing a larger context.
5. Compare an original model and a derivative only with the same benign prompts, temperature, context length, and settings. Measure accuracy and reliability—not merely whether it answers.

As of 20 July 2026, full Kimi K3 weights are announced for 27 July 2026. Before that release and verified conversions, assume repositories claiming to be a complete local Kimi K3 build may be incomplete or misleading.

Official references

• CC Switch v3.17.0 release and provider/local-routing documentation
• Puter Kimi K3 documentation
• Kimi model-selection guide
• ZenMux model list and ZenMux OpenAI-compatible API quick start
• OpenRouter Kimi K3 pricing
• Kimi K3 membership and API pricing and Moonshot AI
• LM Studio model download documentation and offline-operation documentation