Agent tooling

ScaleDown constrained decoding models

ScaleDown hosts models for assigning labels, copying source passages, summarizing documents and trimming the context an agent receives.

Bookmark Keep this in your collection.

Leave a noteWhat would you try with this?

Only you can see your notes.

Source screenshot of ScaleDown constrained decoding models — SOURCE SCREENSHOTFull screenshot ↗

What it does

Classify returns probabilities for labels; Extract gives back source passages with their positions. The cookbooks connect these tools to agent workflows. ScaleDown has not published the family's weights, and its speed and price comparisons are company claims. These models compete with Jev rather than integrate it.

How you can use it

Start with the cookbook for the job you need: labels, exact passages or shorter documents. For extraction, save the original document alongside the returned positions. Your application can then check that each passage appears in the source, rather than trusting a plausible sentence that was never there.