Modelsamos-1-mini
Amos 1 miniSmall enough to run on a laptop CPU. No GPU needed. The API default.
- Size0.5B
- Server time, median27 ms
- Context per pass9K tokens
- Typed decisions, zero-shot64.6%
- Typed decisions, fitted70.3%
- Calibration, KL0.225
- Decision Index 0.3, public17.35
When to use it
- Running on your own machine. About 93 ms a request on a laptop CPU.
- Simple decisions where the answer is in the case.
- Long option lists, such as support intents.
- Telling things apart in photos, such as pet breeds.
Typed decisions
Accuracy on the 2,000-decision test split, zero-shot. We ran each rival on the same 2,000 decisions; Decider 1's score is its makers' own. Frontier language models asked the same questions score between 73 and 78.
- Microsoft-Decision-1closed weights, KL 0.31879.5
- Decider 176.8
- Amos 1 medium76.5
- Amos 1 large76.1
- Perplexity Decider v1.173.9
- Clef72.5
- GPT-6 Luna71.9
- Amos 1 small69.5
- Amos 1 mini64.6
Quotes and photos
- Invoice totals quoted exactly9 / 10
- Points on the named object, RefCOCO50%
- Points on the named object, COCO89%
- Pet breeds93.6%
Limits
- Simple decisions only. It knows little beyond the case.
- Twelve points behind medium on typed decisions.
- Points land on the named object in 50% of RefCOCO photos.
- Reads 9K tokens per pass. English only.
Quick start
Amos 1 mini is the default, so the request leaves out the model field.
curl https://api.amos.scira.ai/v1/systemone \
-H 'Content-Type: application/json' \
-d '{
"state": "Hi, invoice INV-2231 for £306.00 was charged to my card twice this morning. Please refund one of them today.",
"questions": {
"refund": {
"type": "noul",
"instructions": "Is the customer asking for a refund?"
},
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": null,
"technical": null,
"sales": null
}
},
"amount": {
"type": "span",
"instructions": "How much was charged?"
}
},
"model": "amos-1-mini"
}'