Models & ladders

/model with no argument opens a searchable list of everything your endpoints actually serve.

╭──────────────────────────────────────────────────╮
│ search › gemma free    2 of 464                  │
│ ❯   openrouter/google/gemma-4-26b-a4b-it:free      │
│     openrouter/google/gemma-4-31b-it:free        │
╰──────────────────────────────────────────────────╯

Favourites

The catalogue runs into the hundreds and the ones you actually use are a handful. /favourite pins the current model so it rises to the top of /model; running it again unpins it. While the list is open, ctrl+f pins or unpins the highlighted model without leaving, so you can mark several in one pass.

Pinning is independent of default: a favourite is a shortcut to reach a model, not a decision to use it next time.

Switching automatically

Off by default. Turn it on with a ladder of models, largest last. When the conversation outgrows the current model, raunen moves to the next rung rather than failing — and nothing already found is lost in the handover.

{
  "auto_switch": true,
  "fallback": [
    "ollama/qwen3-coder:30b",
    "ollama-cloud/qwen3-coder:480b"
  ]
}

The ladder is yours to define and can mix local and hosted models freely. Adding "free_fallback": true appends every model the providers report as free, roomiest first — see what it costs.

A ladder that remembers

A ladder is only useful if it remembers what just failed. Otherwise a rate-limited model is retried every turn, fails every turn, and each one pays the same tax. The full table of failures and responses is on the cost page; the principles are:

• A model is only remembered once it answers. Escalating updates the saved default, but only after the replacement has actually completed a turn.
• Size only matters when room is the problem. After a refusal, a slightly smaller model that answers beats a larger one that will not.
• A success clears the record. One blip does not suppress a good model.
• Sub-agents share the tracker with their caller, so what one learns spares the other from finding out again.

/status shows what is being held back, which is why a ladder can look shorter than it is:

held back 2 models
          groq/llama-3.3-70b  ·  cooling down for 1m30s
          nvidia/typo         ·  locked out: rejected