Models and Tool Search

Models and Tool Search

Model catalog

When network access and a valid Meta credential are available, the provider refreshes its model list from Meta's Muse model catalog. If the catalog cannot be used, these fallback models remain available:

Model	Input / output / cached input ($/M tokens)
muse-spark-1.2	1.25 / 4.25 / 0.15
muse-spark-1.2-contributor	0.10 / 0.20 / 0.002
muse-spark-1.1	1.25 / 4.25 / 0.15

Fallback definitions provide a 1,048,576-token context window, up to 256K output tokens, image input, and reasoning levels minimal, low, medium, high, and xhigh.

Live catalog metadata can override names, modalities, token limits, pricing, and reasoning variants.

Deferred tool search

Meta models are marked as supporting Pi's deferred tool-loading protocol. This capability does not add a fixed Meta tool catalog and does not turn ordinary requests into web searches.

The flow is:

1. A Pi extension registers tools but initially keeps some inactive.
2. The model calls that extension's loader or activation tool.
3. Pi activates the selected tools and sends their definitions at that point.
4. Meta receives the added definitions through its tool-search protocol without invalidating the earlier prompt prefix.

Tools already active in the session, such as Pi's normal file and shell tools, are not automatically deferred. The exact dynamically loadable tools depend on the installed Pi extensions.

For example, Pi Lens can expose an activation tool that later loads AST search, structural replacement, outlines, LSP navigation, or diagnostic-management tools.

Select a model

pi --list-models meta
pi --provider meta --model muse-spark-1.2

If a refreshed catalog model is hidden by Meta, the provider intentionally omits it from Pi's list.