Skip to main content
An assistant uses one of the models your deployment is configured with. MITHUNAI speaks three provider interfaces, Anthropic, OpenAI and Google, and the operator of your deployment decides which of them it has and where each one points: at the provider’s own service, or at an endpoint that speaks the same interface, such as the Model Gateway. List models returns exactly what your deployment can reach; use a returned id as the assistant’s model.

​
The catalogue

MITHUNAI keeps a catalogue of known models. The list below is the catalogue; your deployment offers only the entries whose provider is configured. A model written as provider/name is accepted even when it is not in the catalogue, as long as its provider is configured; the endpoint behind that provider then decides whether the name is valid. That is how a deployment whose OpenAI interface points at the Model Gateway, or at another OpenAI-compatible endpoint, offers models that are not in this table. A provider with a different interface cannot be configured.

​
Choosing a model for an assistant

  • Set model when you create or update an assistant, or choose it in the console. It applies from the next question; earlier turns in a conversation are unaffected.
  • Leave it out and the deployment picks one it can reach: anthropic/claude-sonnet-5 when Anthropic is configured, otherwise the first chat model in the catalogue that a configured provider serves.
  • A model you name must be reachable. Naming a model whose provider your deployment has no credential for is refused with 400, rather than silently replaced.
  • A model you name must be able to answer. A catalogued embedding model is refused with 400, and List models does not offer one. A provider/name that is not in the catalogue is not checked this way; the endpoint behind that provider decides.
The model changes how answers are written, not what they are grounded in: every model answers from the same retrieved evidence, under the same checks. See How answers work.

​
Embedding models

A collection’s embedding model is fixed when the collection is created, and every passage in it is embedded with that model. Anthropic offers no embedding model, so ingestion needs OpenAI or Google configured. Collections embedded with different models cannot be combined in one assistant; the console shows such a collection as unavailable, with the reason.

​
Who configures providers

Provider credentials belong to the deployment, not to your organisation. If the model list is empty, or a model you need is missing, ask whoever operates your MITHUNAI deployment. See Model gateway for how model calls are retried and what happens when a provider is unavailable.
Last modified on November 7, 2026