VARIANT-1 provides the workspace; you supply model access. Models differ in strengths, context limits, and tool capabilities. The beta release notes will identify the connections tested for that version.
Local models
Local inference uses a supported model file and runtime on your Windows computer. Model weights remain user-supplied, and practical model size depends on the available memory, processing hardware, and runtime support.
This route avoids a cloud token bill for the local model, while still using your hardware, storage, and electricity. A task may separately use online services or browsing, so a local model does not imply that every capability works offline.
API keys and custom endpoints
Use your own provider account and API key, or a supported custom endpoint. Your provider sets its usage prices, quotas, model availability, and data-handling terms.
Check the endpoint and model you intended to select. Model context size, image support, tool use, and other abilities can vary. Start with a small task before relying on a route for substantial work.
Provider-account connections
A supported account connection is another way to authorise access. It still requires the account entitlement, subscription, or other conditions set by that provider.
Signing in does not create a free model allowance. In the initial source preview, Google subscription OAuth also requires an authorized client configuration; one is not bundled.
What you pay for
| Part | Provided by | Cost or requirement |
|---|---|---|
| Windows host | VARIANT-1 | Free download, in preparation. |
| Local inference | Your model and hardware | Supported files/runtime and suitable resources are required. |
| Cloud inference | Your provider arrangement | Provider prices, entitlements, and limits apply. |
No VARIANT subscription, checkout, managed cloud balance, or VARIANT account is required for this beta.
If you reach a limit
Check the error and your selected connection. An exhausted quota, expired sign-in, rejected request, and unavailable local model need different fixes.
Do not assume another route has taken over. If you choose a different model or provider, account for its capabilities and cost before continuing. See Troubleshooting for a practical way to narrow down the issue.