Assigning Model Groups

Cremind never hard-codes a single model. Instead it reaches for a model group — a named slot that points at whichever provider/model you assigned to it. There are two groups:

• high — the reasoning and planning model. This is the heavyweight, more expensive model the agent uses to think through a task and decide what to do next.
• low — the per-tool-call model. This is the cheaper, lighter model used for routine work like the skill classifier and individual tool calls.

Splitting work this way means you do not pay frontier-model rates for every small step — the expensive model plans, the cheap model executes — which gives you lower token cost for the same end-to-end task. For the architectural reasoning behind the split, see Model Groups.

Reading the current assignments

In the web UI, the Model groups section sits at the top of Settings -> LLM Providers. From the CLI:

cremind llm model-groups get

{
  "default_provider": "anthropic",
  "model_groups": {
    "high": "anthropic/claude-opus-4-7",
    "low":  "anthropic/claude-haiku-4-5-20251001"
  }
}

Each assignment is a provider/model pair, where provider matches a NAME from cremind llm providers list and model matches an ID from cremind llm providers models <provider>. The default_provider is the provider used when no explicit provider is requested.

Setting the groups

Set any combination of the three fields; omitted fields keep their current value, and at least one must be supplied.

# Bump the high group to a stronger reasoning model
cremind llm model-groups set --high anthropic/claude-opus-4-7

# Use a fast, cheap model for the low group
cremind llm model-groups set --low anthropic/claude-haiku-4-5-20251001

# Switch the default provider and the high model in one call
cremind llm model-groups set --default-provider openai --high openai/gpt-5

How to choose models for each group

A good starting point:

• Put your strongest available reasoning model in high. This is where quality of planning matters most.
• Put a fast, inexpensive model in low. Routine tool calls and the skill classifier do not need a frontier model, and using a cheap one here is the main source of the lower token cost.

You can mix providers — for example, a high group on one provider and a low group on another — as long as both providers are configured and their models have been discovered.

Where the assignments are stored, and precedence

Model-group assignments are stored in SQLite on the server. These stored values override the corresponding settings in the TOML settings.toml file. In other words, TOML provides the baseline a fresh server starts from, but once you set a group through the CLI or the web UI, the database value wins.

Per-tool overrides

The two groups are the default the agent reaches for, but an individual tool can pin which group it should use through its default_model_group configuration. This lets a specific tool always run on the high group (or always on low) regardless of the agent's normal choice for that step — useful when one tool needs more reasoning headroom than the rest.

Next steps