Model Configuration
Overview
The Model Configuration page is used to manage LLM providers and model parameters. The App side supports configuring multiple providers simultaneously, and the system automatically switches the API format and capability detection based on the selected model.
Supported Providers
| Provider | API Format | Description |
|---|---|---|
| Anthropic | anthropic | Claude series models |
| OpenAI | openai | GPT / o series models |
| DeepSeek | openai | DeepSeek-chat / reasoner |
| Qwen (Tongyi Qianwen) | openai | Qwen-plus / Qwen-max |
| Gemini | openai | Google Gemini series |
| Custom | openai / anthropic | Third-party services compatible with the OpenAI or Anthropic format |
Configuration Items
| Parameter | Description |
|---|---|
| Provider Selection | Select an LLM provider from the drop-down list |
| API Key | Provider API key (stored encrypted) |
| Base URL | API endpoint address (required for custom providers) |
| apiFormat | API protocol format: anthropic or openai |
| Model Selection | Select from the provider's list of available models |
| tokensRespLimit | Maximum number of output tokens per turn |
Model Capability Flags
The system automatically detects model capabilities via resolveCapabilities, which affects multimodal input processing:
| Capability | Description | Impact |
|---|---|---|
| vision | Image understanding | Whether image input is accepted |
| document | Document understanding | Whether document input such as PDF is accepted |
| audio | Audio understanding | Whether audio input is accepted |
| tool_use | Tool calling | Whether function calling is supported |
Context Window
Default context window size for each model:
| Model | Context Window |
|---|---|
| Claude Sonnet 4 / Opus 4 / Claude 4 | 200,000 tokens |
| GPT-4o / GPT-4o-mini / GPT-4-turbo | 128,000 tokens |
| GPT-4.1 | 1,000,000 tokens |
| o3 / o4-mini | 200,000 tokens |
| DeepSeek-chat / DeepSeek-reasoner | 64,000 tokens |
| Qwen-plus / Qwen-max | 128,000 tokens |
For the complete list of models, refer to Supported Models List.
Operation Guide
How to Configure a Model
- Open the App → left-side settings menu → Model
- Select the LLM provider you want to use from the provider list (e.g., Anthropic)
- Enter the API Key provided by that provider
- (Optional) If you are using a third-party compatible service, change the Base URL to the endpoint address provided by the service
- Select a specific model from the model drop-down list (e.g., Claude Sonnet 4)
- Click Save
Note: The API Key is stored locally in encrypted form and is not uploaded to the cloud. Never share your API Key with others or commit it to a code repository.
How to Switch Models
After configuring multiple providers, you can switch the currently used model at any time. After switching, the next conversation will use the new model. The history of existing conversations is not affected.
Model Selection Recommendations
Different scenarios call for different models. The following are recommended combinations:
| Use Case | Recommended Model | Reason |
|---|---|---|
| Everyday conversation and lightweight tasks | GPT-4o-mini / Qwen-plus | Fast response, low cost |
| Complex reasoning and long document analysis | Claude Opus 4 / Claude Sonnet 4 | Strong reasoning, 200K long context, supports document input |
| Code development and debugging | DeepSeek-chat / Claude Sonnet 4 | Outstanding code comprehension |
| Ultra-long context (entire codebase) | GPT-4.1 | 1M token context window |
| Requires image understanding | Claude Sonnet 4 / GPT-4o | Supports vision capability |
| Cost-sensitive | Qwen-plus / DeepSeek-chat | Lowest price |
| Requires deep reasoning (math/logic) | o3 / DeepSeek-reasoner | Reasoning-chain models, but DeepSeek-reasoner does not support tool calling |
Tip: If you are unsure which to choose, we recommend starting with Claude Sonnet 4 — it delivers balanced performance across reasoning, code, tool calling, and multimodality.
Real Impact of the Context Window on Users
The context window determines how much content the Agent can "remember" within a single conversation:
| Window Size | Roughly Equivalent To | Applicable Scenarios |
|---|---|---|
| 64K tokens | ~100 pages of plain text or ~5 medium-sized code files | Brief conversations, single-file operations |
| 128K tokens | ~200 pages of plain text or ~10 code files | Medium-complexity projects, multi-file operations |
| 200K tokens | ~300 pages of plain text or ~15 code files | Complex projects, long conversations, PDF document analysis |
| 1M tokens | ~1500 pages of plain text or ~70 code files | Very large codebase analysis, book-level documents |
When the conversation content approaches the window limit, the system automatically triggers Context Compression, intelligently retaining the most important information.
Impact of tokensRespLimit
tokensRespLimit controls the maximum number of tokens in the Agent's reply per turn:
| Setting Value | Impact |
|---|---|
| Too small (e.g., 500) | The Agent's reply may be truncated, and complex analysis or long code generation will be incomplete |
| Moderate (e.g., 4096, recommended default) | Balances reply completeness and response speed |
| Too large (e.g., 32000) | The Agent can produce very long replies, but this increases wait time and cost |
Recommendation: Keeping the default value is sufficient. If you frequently encounter truncated replies, you can increase it appropriately.
FAQ
Invalid API Key
- Confirm the Key was copied correctly (watch for leading and trailing spaces)
- Confirm that the correct provider is selected for the Key
- Confirm the Key has not expired or been revoked
Request Timeout
- Check your network connection
- If using a custom Base URL, confirm the address is reachable
- Some models may respond slowly under high load; you can retry later
Poor Reply Quality
- Try switching to a more powerful model (e.g., from GPT-4o-mini to Claude Sonnet 4)
- Check whether tokensRespLimit is set too small
- Optimize your prompt and provide clearer instructions
Related Documents
- Supported Models List — Complete model capability matrix
- 5-Layer Context Compression — Context window management strategy