# Base models

Every Model starts from an open base model, and Facet builds your version on top of it. Facet picks the base when it drafts the Model: the default, or the smaller one when it can meet every eval's bar. The draft's review says which, and you can change it there.

| Base model | Best for | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|---|
| Qwen3.5 0.8B | Simple, high-volume tasks: short labels, routing, yes or no | $0.04 | $0.12 |
| Qwen3.5 4B | Most tasks: extraction, rewriting, judgment calls (the default) | $0.08 | $0.25 |


Both are Qwen3.5, from Alibaba's Qwen team, under the Apache 2.0 license, so an exported Model runs anywhere. See [Limits](/docs/limits) for the longest input each takes.

## Which to pick

- **Keep Facet's pick** unless you have a reason. The default handles most Models well.
- **Try the smaller one** when answers are short and the task is simple, such as a label from a short list, and you make many calls. It costs about half as much per token.
- **Try both in one go.** Ask for it when you describe the task or a fix, such as "try both bases". Facet trains each on the same examples and scores them on the same evals, cheapest per call first, and stops at the first that meets every bar. When two score within the noise, the cheaper per call goes live. The quote prices every run.

## What stays the same

A version records the exact revision of its base, so it answers the same way until you change it. Facet emails a workspace's Owners at least 90 days before it retires a base its Models use. After that, Models on it keep their history and exports but stop answering calls; to move one, ask a Change to try a current base.

## Next steps

- [Pricing](/docs/pricing): what building and calling a Model costs
- [Export your model](/docs/export): run a Model on your own hardware