English | 中文
DeepSeekGUI uses the Harness model and credential services directly. A model configured in Settings → Models is available to DeepSeekGUI without a second provider file or credential store.
- Open Settings → Models.
- Select the DeepSeek provider.
- Enter your DeepSeek API key.
- Fetch or refresh the model list when the page offers that action.
- Select the model you want to use for the session.
The provider may reject a key that is missing, invalid, expired, or not authorized for the selected model. DeepSeekGUI reports that failure without copying the key into its diagnostics.
The v1.1.1 catalog includes DeepSeek V4.1 Flash (deepseek-flash) with image input. Select an image-capable model available to your account, attach the image, and send a clear instruction; configure custom providers to match their actual capabilities.
An image-capable model can inspect screenshots, diagrams, UI states, and other visual inputs. A text-only model cannot infer an attachment's contents; switch models before asking it to reason about the image.
Use the custom-provider form when you need an OpenAI-compatible endpoint or another provider exposed by the Harness settings page. Enter only the values required by that provider:
- Provider name and base URL.
- API key or credential, when required.
- Model id.
- Input capabilities, including image support when the endpoint accepts images.
Do not mark a text-only endpoint as image-capable. The selector uses this metadata to decide whether attachments can reach the model.
The active model belongs to the Harness session. Change it from the model selector in the conversation interface. Existing messages remain in the session; the next model receives the context that Harness prepares for the next turn.
Refresh the provider's models. For a custom endpoint, add the exact model id supplied by the provider. Availability can differ by account and region.
Return to Settings → Models and store the key for the same provider route the session uses. A key stored for one custom provider does not authorize another route.
Confirm that the selected model advertises image input and that the endpoint itself accepts OpenAI-compatible image_url content. Then start a small test message with one image and one direct question.