Models Overview
Xagent supports multiple types of AI models, each serving a different purpose in task execution. You connect providers under Models, then assign models to roles and set defaults.
Model Types
Internally each model is bound to a configuration role (config_type): general, small_fast, compact, visual, or embedding.
| Role | config_type | Purpose |
|---|---|---|
| Main Model | general | Required. Powers planning, reasoning, and decision-making for all task execution. |
| Fast Model | small_fast | Optional. Used for steps the planner marks as simple — lowers cost and improves speed. |
| Long Context Model | compact | Optional. Compresses conversation history when it exceeds the token threshold (default 32,000). |
| Vision Model | visual | Optional. Required for image-understanding tools (understand_images, describe_images, detect_objects). |
| Embedding Model | embedding | Converts text to vectors for knowledge base search and RAG. |
Built-in Models vs. Your Own
There are two kinds of model you can use, and the difference matters for cost.
| Built-in model | Your own model | |
|---|---|---|
| Who supplies the API key | Xagent — nothing to set up | You do, when adding the model |
| How it is paid for | Deducted from your team's monthly AI credits | Billed to you directly by the provider |
| Uses AI credits? | Yes | No, never |
| Can you edit or delete it? | No — shared and read-only | Yes, it is yours |
Built-in models appear in your model list with a Built-in badge. You can choose whether to use them, but you cannot change or remove them.
Available built-in models
| Model | Provider | Context window | Good for |
|---|---|---|---|
deepseek-v4-flash | DeepSeek | 1,000,000 tokens | General work. This is the built-in default for the Main role. |
qwen3.6-plus | Alibaba DashScope | 1,000,000 tokens | Understands images as well as text. The built-in default for the Vision and Long Context roles. |
If you never choose a model yourself, your agents use these built-in defaults.
AI Credits
AI credits are the allowance your team spends when it runs a built-in model. Each plan includes a monthly amount that resets each billing period, and roughly 1,000 credits is about one US dollar of model usage. Bringing your own model key bypasses credits entirely.
A model named the same as a built-in one is still yours
Credit charges are matched to the specific built-in model, not to its name. If you add your own model and happen to call it qwen3.6-plus, it is still your model on your key — it does not consume AI credits.
Model Sharing
Beyond the built-in models, admins can configure additional shared models for everyone in the workspace, set permissions, and monitor usage. Regular users can then use those shared models without managing their own API keys.
Admin configured
Model sharing is set up by admin users in the Models settings. Regular users can only access shared models.
Configuration
- Add a provider — OpenAI & OpenAI-compatible, Anthropic, Google, Xinference, or a custom endpoint.
- Configure models — select the model, enter credentials, set parameters, and test the connection.
- Set defaults — choose default models globally, per task, or per agent.
Requirements
- Required: at least one LLM as the Main model (Claude 4.6 Sonnet or higher recommended).
- Knowledge Base: an embedding model.
- Image analysis: a vision model.
- Cost optimization: a fast model.
- Long conversations: a long-context model.