Auto chooses a model based on the role and run mode. Pin a model when you need repeatability, a specific context window, or a provider required by your organization.

Open Settings > Models and keys > On-device AI. Choose Qwen3 Small (0.6B, about 350 MB), Balanced (1.7B, about 1 GB), or Larger (4B, about 2.3 GB). Codelit checks browser GPU support, processing capacity, reported memory and available browser storage. A suggested choice is a starting point; browsers cannot reliably report available GPU memory.

1. 1Choose Load model to download and test that model. Opening settings does not download model weights.
2. 2Wait for Ready on this device and choose Use this model if it is not already selected. In use on this device confirms your selection. Local AI supports short text drafts, summaries and rewrites, with no per-token AI fee.
3. 3Unload before choosing a different size. The download stays cached, but reloading the page or changing accounts clears readiness.

Open Settings > Models and keys, then add an OpenRouter, OpenAI, Anthropic, Gemini, Cerebras, Groq, or Mistral key under Bring your own API keys. Codelit saves the key in this browser and refreshes the model list when the provider allows discovery. Choose a model and run a request to check access; saving a key alone does not verify it. OpenCode Zen free access is paused in Codelit until external eligibility is verified. Existing Zen keys can be removed in Settings. Choose another model explicitly before running a saved Zen selection.

After saving, the key input is cleared and Settings shows only the provider and final characters. Choose Replace to enter a new secret or Clear to remove that provider. Codelit never remounts the saved value into the page or sends it to password-manager autofill.

The run cost control shows a range before execution and a line-item breakdown from the same estimator used in History. Dry run has no model cost, while user-funded modes charge the user's provider account.