mukaddamzaid/amos-1-large

Amos 1 large

A 27B decision model. Give it a case (text, JSON or a photo) and a schema of typed questions; it answers every question in one forward pass, each with a probability for every answer you allow. It writes no text, so there is nothing to parse and no malformed output. Our most capable model, for the hardest calls: questions that take several steps, long policies and trade-offs, and cases that need broad knowledge.

Part of the Amos 1 family: large · medium · small · mini. Playground · API docs · Model page

Results

Typed decisions, test split: 2,000 decisions from four workflows it never trained on, zero-shot.

	Accuracy	KL ↓	Brier ↓
Amos 1 large	76.1%	0.126	0.069
Microsoft-Decision-1 (closed weights)	79.5%	0.318
Decider 1 (its makers' number)	76.8%	0.096
Perplexity Decider v1.1	73.9%	0.466
Clef (Cloudflare)	72.5%	0.191

Rivals run by us on the same 2,000 decisions, except Decider 1, whose number is its makers' own.

With each question's priors fitted on the benchmark's own train split and applied at answer time (the weights do not change), as Decider 1 also reports: 80.25% (Decider 1: 80.1%).

Decision Bench, all 1,071 real records, images included, run with the benchmark's own harness: 94.9%. The leaderboard lists Perplexity Decider v1.1 at 94.5%, DeepSeek V4.1 Flash and Gemini 3.5 Flash at 94.2%, Jev 1.13 at 93.8%.

Decision Index 0.3, public part, with the board's own kit: 65.11. By area: knowledge 52.6, language 72.3, retrieval 66.2, tools 79.7, arts 50.0. The board's ranking adds private tests that only its maintainers run.

Other benchmarks, as served
JevBench public easy / standard / hard	100.0 / 100.0 / 84.7
Ten public benchmarks (RAGTruth, JudgeBench, Belebele, WinoGrande, ContractNLI, TruthfulQA, Circa, Financial PhraseBank, TabFact, BBH)	89.8%
AG News / DAIR Emotion / Banking77 (77 labels)	89.1 / 58.5 / 90.5
Invoice totals quoted exactly	10 / 10
Box centred on the named object, RefCOCO	91.9%
Box covering at least half the object, RefCOCO	71.3%
Box centred on the named object, COCO	98.4%
Our copy of the Decision Index vision board's public tasks (8 tasks, board weights)	76.69
BLINK, chance-corrected	65.9
Reranking the BM25 top 30, nDCG@10, mean of SciFact, NFCorpus and FiQA	0.517
Server time, median, three-question request	122 ms

Use

Hosted, no key: POST https://api.amos.scira.ai/v1/systemone with "model": "amos-1-large".

Or run it yourself. The serving package is in serve/ (Python 3.12, PyTorch 2.14; versions pinned to the ones these numbers were measured with):

hf download mukaddamzaid/amos-1-large --local-dir amos-1-large
pip install ./amos-1-large/serve
python -m amos.serve amos-1-large --device cuda --port 8000 --max-len 131072
curl http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Invoice INV-2231 for £306.00 was charged to my card twice. Please refund one.",
  "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}}
  }
}'

Question types: yes/no (noul), pick one (choice, up to 255 options), rate (score), pick any (multi), quote (span) and point (region, with a photo in images). It runs on an 80 GB GPU and reads up to 131,072 tokens per request.

File	What
model/ (weights, tokenizer, image processor, heads.pt)	The model
calibration.json	Quote decoding settings the server reads
serve/	The serving package (amos.serve)
typed_decisions.json	The typed-decisions test scores above

Limits

• The slowest of the four.
• Long quotes, such as a whole clause or address, can come back cut short.
• English only.