Custom models

A custom model is a model that Intric doesn’t provide, which you connect yourself. It might be a model you host on your own infrastructure, a private deployment with a cloud provider, or a vendor that isn’t in the catalogue. After you connect it, people pick it the same way as any other model, and the Space’s security classification decides where it can be used.

Before you start

• Custom models must be turned on for your organization. If there’s no Add custom button in the Add a model dialog, your organization doesn’t have the feature yet. Contact Intric to have it enabled. Only Admins can create custom models.
• The endpoint must speak the OpenAI Chat Completions API. Most self-hosting servers and gateways, such as vLLM, Ollama and LiteLLM, offer it.
• Have three values from your provider ready: the base URL, the exact model name, and an API key if the endpoint needs one.

Adding a custom model

Add a custom model

1. Go to Admin → Platform → Models, select Add model, then Add custom.
2. Name your model. The nickname is what people see wherever they pick a model, so use a name they’ll recognise. A preview shows how it will look in the list.
3. Connection. Enter the Base URL, the Model name and the API Key.
4. Hosting location. Choose the country or region where the model runs.
5. Context window. Enter how many tokens the model can hold at once. You’ll find this in your provider’s model documentation.
6. Select Create model.

Intric creates the model, adds it to your organization straight away and runs a connection test. The panel then switches to the model’s full settings, so you can finish configuring it without reopening it.

The connection fields

• Base URL — the address Intric sends requests to, ending at the API version, for example https://api.your-domain.com/v1. Intric adds the rest of the path itself, so don’t include /chat/completions.
• Model name — the exact model identifier the endpoint expects in each request, for example vendor/model-latest. This is often different from the nickname.
• API Key — sent as an Authorization header, with Bearer added in front automatically. After you save a key, it’s never shown again. Leave the field empty to keep the saved key, or enter a new one to replace it.

Testing the connection

The test sends one short, real request to the endpoint using the saved settings. It shows Success! or Connection failed, along with the error the endpoint returned. If you change the connection fields later, Save & test saves them and then runs the test again.

A successful test confirms that Intric can reach the endpoint and get a response. It doesn’t tell you anything about the quality of the answers.

What to configure after creating it

For a model from the catalogue, Intric already knows the price, the capabilities and where it runs. For a custom model it doesn’t, so fill these in yourself:

• Description — the tooltip people see when they hover over the model name in a Space or assistant, and on the model card.
• Features — the context window, Vision support if the model accepts images, and Enable tools if it supports tool calls. Turn tools on only if the endpoint really handles tool calls, because web search, integrations and MCP tools all depend on them.
• Extra parameters — a JSON object forwarded unchanged with every request, such as {"temperature": 0.2} or a reasoning setting your server understands. Leave it empty to send nothing extra. Intric sets the keys model, messages, stream, stream_options, tools, tool_choice and n itself, so you can’t use them here.
• Token cost — the input and output price per million tokens, and the currency. See Token cost for custom models below.
• Performance — the model’s speed and intelligence ratings. See Ratings for custom models below.
• Hosting location — shown next to the model, so people can see where their data goes. It also matters for classifications that limit processing to certain regions. See Processing regions.
• Security classification — which Spaces may use the model. Set the highest level the model is approved for, and it’s automatically available at every lower level too. See Model classification.

Ratings for custom models

A custom completion model can be bound to a benchmark Intric has synced, so its speed and intelligence ratings follow that benchmark instead of figures typed in by hand — and stay right as the benchmark is updated. Intric suggests a matching benchmark from the model’s name when you add it; accept the suggestion, pick a different benchmark, or enter the figures manually.

Token cost for custom models

Important

For standard vendor models, Intric already knows the pricing and applies it automatically to usage and cost reporting. For custom models — self-hosted or organization-specific deployments — Intric has no way of knowing the price, so set the token cost yourself on the model’s configuration, matching your actual negotiated rate rather than a public list price. Set both rates to 0 if the model costs nothing per token. Keep it current: an outdated or missing value leads to under- or over-reported figures wherever that model’s cost is shown, including Usage and Billing.

How custom models differ from catalogue models

• Curated picks never choose a custom model. An assistant set to Balanced, Fast, Economical or Smart always runs on a catalogue model. To use a custom model, choose it by name.
• Custom models aren’t tied to a sub-processor. Intric can’t know who operates your endpoint, so it doesn’t gate the model through Sub-processors. Record the operator in your own DPA documentation, and use the hosting location and security classification to control where the model may be used.
• Add custom creates a completion model. Custom embedding, transcription and image models aren’t created from this dialog.

Editing and removing a custom model

Open a custom model’s actions menu on the Models page and choose Edit model to change any of its settings, or Remove to delete it.

Important

Removing a custom model deletes it permanently. Unlike a catalogue model, it doesn’t go back to the catalogue. Intric won’t delete a custom model that assistants, personal assistants, apps or conversations still use, so move those to another model first.