Skip to content
Build Perch
Browse documentation

Choosing a model

Short answer: The model picker is one searchable, sortable list: every connected provider’s models together, each row showing its context window, its real per-token rate, and whether it supports tools, reads images, accepts audio, and is offered under zero data retention. A filter bar narrows the list by the same traits, by a minimum context window, and by rate; the zero-data-retention filter starts on whenever that setting is on, and the list follows the setting if you change it while the picker is open. Each chat is pinned to one model but you can switch at any point without losing history.

What is the model picker?

Every chat picks its model when it’s created, from a single list that combines whatever providers you’ve connected: OpenRouter’s live catalog, your local Ollama models, and Ollama Cloud’s hosted models together. Rather than a curated “recommended” shortlist, the list is fully sortable by name, context window, price, or tool support, and searchable by name or id. Your starred favorites and the 10 most recently added OpenRouter models sit at the top as discovery aids; everything else groups by provider underneath.

Media models (image, speech, music, video generators) aren’t part of this same chat-model list. They’re picked separately, per category, from the relevant row in a chat’s configuration, because OpenRouter is the only provider that serves media models. That pick is independent of your chat model: whichever provider the conversation runs on, media still generates through OpenRouter.

How do I pick or switch a model?

  1. Open Connect AI Model (new chat) or Switch Model (existing chat).
  2. On the OpenRouter tab, decide first whether to pick at all. The Let OpenRouter choose section above the list holds OpenRouter’s routers: pick one and OpenRouter chooses a model for each message, and each reply shows which model answered. The Auto Router takes a spend band (a price range, so the top band excludes cheap models as firmly as the bottom one excludes expensive ones), the Pareto Code Router a coding tier, and the Free Models Router picks from free models with their rate limits and privacy caveats. The band or tier is saved with the chat and can be changed later through Switch Model.
  3. Otherwise search by name, or browse the Favorites / Recently Added / per-provider groups. Star a model to pin it to Favorites for next time.
  4. Sort the list by clicking a column header: Model, Context, Rate (in/out), or Tools. Every section reorders together. For Ollama Cloud the rate column is Size instead (the model’s approximate parameter count), since a subscription has no per-token rate.
  5. Compare rows directly: each shows the model’s context window, its $/1M-token input/output rate (or “Free” / “See provider” when a clean rate doesn’t apply), and a checkmark if it supports tool calls. Chat models without tool support carry a “No Tools” badge.
  6. Pick a model and connect. Switching an existing chat’s model preserves its history; a banner warns you first if the new model can’t do something the current one can (no tools, no vision, no audio, or a much smaller context window than what’s already in the conversation).

Good to know

  • Every rate links to the model’s page on OpenRouter or Ollama for the full pricing breakdown. There’s no in-app spend total; Build Perch shows per-model rates only, and you pay the provider directly.
  • Media generation is a separate pick. Image, speech, music, and video models are chosen from their own row in a chat’s configuration, pinned to that one category, and only OpenRouter models appear there because only OpenRouter serves them. This is independent of the chat model: an Ollama or Ollama Cloud chat still generates media, routed to OpenRouter, once you’ve configured an OpenRouter key.
  • Switching models mid-chat is designed for cost-tiering a workflow: run a cheap model for the legwork, then switch the same chat to a frontier model for the final pass. History carries over either way.
  • A model without tools can still receive a mode with tools assigned; the switch banner flags it and offers to fall back that chat to Chat Only automatically on confirm, rather than leaving tools silently broken.
  • Local Ollama models show live capability data (size, quantization, context window, tool/vision support) pulled from Ollama itself, so you’re comparing your actually-installed models on the same terms as the remote catalog.