Nimbus is built around open models, but open model lists can be overwhelming. A model can be famous, huge, tiny, text-only, vision-capable, audio-only, cloud-hosted, local-only, converted, unconverted, or simply wrong for your device. Recent picker updates make those differences more visible before you spend time and storage on a download.
Start from the module
The easiest path is still to open the module first. Gale shows chat models. Mirage shows image models. Overture shows speech models. Stratus shows vision models. The picker filters around what that module can use, so you are not starting from the entire internet.
When a model can work in more than one place, Nimbus uses capability hints to route it. A vision-language model can appear for Stratus or Gale image attachments. A speech model can be offered for Overture instead of mixed into chat.
Read the chips before installing
Model rows now do more work. They show install state, estimated storage, runtime family, and compatibility hints. If a model needs conversion on Mac, is not supported on the current device, or is better suited for another module, Nimbus tries to say that in the row instead of letting you find out after the download.
Auto still respects your choice
Auto routing is useful when you do not care which exact model answers. But if you have chosen a Gale, Mist, Mirage, or Overture model yourself, Nimbus now tries harder to keep the active route aligned with that selection. The model chip and the actual mounted model should tell the same story.
How to pick a model without overthinking it
- Use the curated list first if you just want something that works.
- Check the storage estimate before downloading large image or vision models.
- Prefer smaller models on iPhone if you care about speed and battery.
- Use Advanced only when you knowingly want unsupported, experimental, or conversion-required models.
- If a model does not appear in a module picker, it probably cannot serve that workflow yet.
The model picker is becoming less of a file browser and more of a setup assistant. That matters because local AI should not require users to memorize file formats just to ask a question, dictate a note, or make an image.
Tagged Models · Published Jun 12, 2026 · Browse the model catalog · Back to all posts