ModelsFour sizes, one API.
Every model reads the same request and answers in the same shape. Name one with the model field; leave it out and Amos 1 mini answers. Larger models take the harder calls, smaller ones the volume.
- 27B · 128K context
Amos 1 large
Long policies and calls that take several steps.
typed 76.1% · 122 ms
- 9B · 128K context
Amos 1 medium
Most production decisions, in half large's time.
typed 76.5% · 61 ms
- 2B · 256K context
Amos 1 small
High volume, with photos, quotes and boxes.
typed 69.5% · 48 ms
- 0.5B · 9K context
Amos 1 minidefault
Runs on a laptop CPU, with no GPU needed.
typed 64.6% · 27 ms
Side by side
Measured as served, on data none of them trained on. Typed decisions is its 2,000-decision test split; the Decision Index is the public part of edition 0.3, run with the board's own kit on every row.
Typed decisionsaccuracy, %
- Amos 1 large76.1
- Amos 1 medium76.5
- Amos 1 small69.5
- Amos 1 mini64.6
CalibrationKL to the gold, lower is better
- Amos 1 large0.126
- Amos 1 medium0.151
- Amos 1 small0.181
- Amos 1 mini0.225
Decision Index0.3, public part
- Amos 1 large65.11
- Amos 1 medium52.55
- Amos 1 small31.92
- Amos 1 mini17.35
Server timemedian, three questions
- Amos 1 large122 ms
- Amos 1 medium61 ms
- Amos 1 small48 ms
- Amos 1 mini27 ms
How they compare with other decision models is on the index.