Models
The Models page lets you view and manage the models that Agents can use, including system-hosted models and company-level and personal model configurations.
Version: 6.1. Intended for: model users and maintainers with model configuration permissions.
Search keywords: Model management, models, system models, company models, personal models.
1. What model management solves
Model management helps you connect the large models you need, set their availability scope and default selection, and view your own usage. Individuals can choose models that suit their tasks, and admins can centrally maintain company models and automatic model selection rules.
Models handle understanding, reasoning, and generation. Agents combine models with instructions, tools, and knowledge into concrete working capabilities. Connecting a model doesn't mean that you've created an Agent.
| Model type | Use cases | Who can maintain it |
|---|---|---|
| System models (hosted) | Uses models provided by the platform without filling in connection information yourself. Availability is limited by the company configuration. | Users set their own enable/disable preferences, and admins control the company-wide usage scope. They aren't added or deleted through the personal or company model forms. |
| Company models | Connected centrally by the company for members of the company to use within the allowed scope. | Admins with model management permissions maintain the connection, capabilities, and company-level enable/disable. Members set their own usage preferences. |
| Personal models | Models that you connect for your own use, visible and usable only to you. | Created, edited, enabled/disabled, and deleted by yourself. |
A visible model isn't necessarily callable in every scenario. The company switch, your own enable/disable preference, your account's model usage limits, and the Agent's model configuration all affect final availability. Models belong to the Capability Center, so you don't need to connect them again from a workspace.
2. Connect a personal or company model
Connect personal models with + Personal Model in Capability Center > Models. To connect company models, admins go to System Admin > Model Management and click + Add Company Model. Regular members don't maintain company connection information in the Capability Center.
- Enter the model ID, Base URL, and API Key. The model ID is the model name that the provider actually receives. Enter the Base URL as the provider's compatible API address, and check the address format against the form hints. The API Key is required when you create a model.
- Fill in the display name, provider, and description as needed. If the display name is left empty, the model ID is used. After you use Quick Fill, you still need to check the selected provider and the actual connection information.
- Expand Advanced settings and fill them in based on the capabilities and limits that the model actually supports. For details, see the next chapter.
- Click Connection test, and confirm the connection or save only after it shows that the connection succeeded. If you modify the model ID, provider, Base URL, or API Key, you need to test again.
- After saving, check the name and enabled status in the model list. If you're told that the name is a duplicate, go back and change the name, or confirm the automatic rename. Connecting a model with the same name doesn't overwrite the existing model.
The connection test only verifies that the basic connection information works. It doesn't verify image input, reasoning mode, maximum context, or maximum output. After you save the configuration, check the results with a small task that matches the actual use case.
3. Configure model capabilities and runtime limits
Personal and company models use the same advanced setting fields. Base them on the actual capabilities of the connected model. Don't try to "add" capabilities that the model doesn't have by increasing values or turning on options.
| Config items | Purpose | What to enter |
|---|---|---|
| Auto-compaction threshold (tokens) | Auto-compaction is triggered when the conversation history reaches this threshold. | A positive integer no greater than the model context limit. It isn't the model's maximum capacity. |
| Model context limit (tokens) | The total context capacity for input and output that the model can accept. | A positive integer, based on the provider's actual limit for the current model. |
| Maximum output (tokens) | Limits the output length of a single model response. | A positive integer no greater than the model context limit. The product default is 32,000. Lower it manually if the provider supports a smaller limit. |
| Supports image input | Declares that the model can receive image content. | Turn it on only if the model really supports it. The basic connection test doesn't upload images to verify this. |
| Reasoning mode | Select Follow default, on, or off for reasoning. | Only expresses the reasoning intent, without setting the thinking effort or token budget. The actual effect still depends on support from the model and provider. |
If the connection succeeds but a context or output limit error occurs at runtime, first check the configuration here against the provider's limits. Changing only the capability fields doesn't require a new basic connection test, but you still need to verify the actual runtime results of the corresponding capabilities.
4. Set default models and Company Auto
4.1 Set personal and company defaults
In Personal default under Capability Center > Models, select a currently available system, company, or personal model and save. When Company Auto is available, you can also select it as your personal default. The personal default is used first when you create a default Agent conversation, and can also be referenced dynamically by the company default and Agent configurations.
Clicking Clear only removes the personal default setting and doesn't delete the model. When the existing default becomes unavailable, the page keeps showing a notice. Select a new one or clear it. Don't assume that it can still run just because the old name is still shown.
Admins edit the company default models in System Admin > Model Management. For each business purpose, they select an available system model, company model, or Company Auto, or choose to dynamically reference Personal default. When the personal default is referenced, the current executing user's setting is used, not the admin's own personal model.
Changing the defaults doesn't force all Agents and existing conversations to switch to the new model. A fixed model still points to a specific model. Only when Personal default or Company default is selected is the model resolved from the corresponding default configuration, and availability requirements still apply.
4.2 Use and maintain Company Auto
Company Auto automatically selects a model based on the rules configured by admins and the task type. It isn't a new provider that requires users to fill in connection information. Users can enable their own Company Auto preference on the Models page and select it wherever model selection is allowed. If it shows as unavailable, contact an admin to check the company configuration.
Company Auto is currently a Beta feature and may not hit the cache. Before you turn it on, check the results and token usage with actual tasks.
Admins configure rules in the Company Auto area of the same model management page: first create a default rule, enter the rule name, set the models used for the classifier, simple tasks, regular questions, and complex questions, and configure the classifier timeout, failure fallback, and target model exception strategy. If images need to be processed, also set an image model. Rules reference system or company models available to the company and don't use personal models.
The default rule applies to members who aren't assigned a special rule. When you need differentiated configurations, add a special rule and select the members it applies to. After saving, check the enabled status of the rules and Company Auto. If a member already belongs to another special rule, you need to confirm before moving them.
Test classification only calls the classifier model. It doesn't call the target model to complete the task and can't replace actual verification of the target model. If images need to be processed, configure a separate image model that is available and supports image input. If the company default still references Company Auto, change the default reference first, and then disable Company Auto.
5. Use in Agents, workspaces, and conversations
Select a model for an Agent. In the Agent configuration, select a fixed model or a default reference based on its purpose. A fixed model suits cases where you want an explicit runtime choice. A default reference suits cases where the model should follow personal or company configuration changes. Whether users can switch models in a conversation is determined by the Agent's configuration. For full settings, see Agent.
Use in conversations. In a conversation that allows model switching, select a currently available option from the model selector and then send the task. If a model is unavailable or restricted by the Agent, select another one as prompted, or contact the Agent maintainer. For sending, continuing conversations, and viewing results, see Conversations.
Use in workspaces. After you enter a space, use models in the conversations within the space. A shared space doesn't automatically make one member's personal models available to other members. When a team needs to share a model, an admin connects it as a company model. For space types, member permissions, and common capability settings, see Create a space.
6. Enable/disable, usage, and FAQ
6.1 Personal enable/disable vs. company-level control
Toggling a model switch in the Capability Center only changes whether you use it and doesn't change other members' configurations. When admins need to change the company-wide usage scope, they use the company-level switch in System Admin. A model that the company has disabled can't be restored with a personal switch.
After an admin turns off a company model, neither members nor admins can keep selecting it on the model usage page or in conversations. Admins can still find the configuration on the management page and enable it again. System models are also limited by whether the company has turned on the system model service, and the switch of a single model can't override this limit.
6.2 Edit and delete models
Personal models are maintained by their owners, and company models are maintained by admins with the required permissions. When editing, leaving the API Key empty keeps the existing key. After you replace the key or modify other connection information, run the connection test again and then save.
Before deleting a model, check whether default settings, Agents, or Company Auto rules depend on it, and arrange a replacement. If you're told that the model is still used by the company default configuration, change the corresponding default first and then delete it. After deletion, the model is removed from the available list. Don't delete models as a way to clean up conversation history or usage records.
6.3 View usage and troubleshoot
In Usage Stats on the Models page, select Last 7 days, Last 30 days, or the last 1 year, and filter by model to view your input and output tokens, number of conversations, and trends. These are personal usage statistics. For the organization-wide entry and permissions, see System Admin.
Connection test fails: Check the model ID, compatible API address, API Key, and provider availability. Correct them according to the error message and test again. Don't keep saving the same failed configuration.
Model can't be found in the list or can't be used: First confirm the model type, your own enable/disable status, and usage permissions, and then ask an admin to check the company switch and account limits. If only a specific Agent can't use it, also check that Agent's fixed model and switching restrictions.
Company Auto or default model is unavailable: Select another available model, or ask an admin to check the default references, Auto rules, and their target models. A successful basic connection, a successfully saved rule, and a successful classification test don't mean that all tasks have been verified to run.