Amos 1 large

Amos 1 large

Answers typed questions about a case, given as text, JSON or a photo, in one pass with a probability for each allowed answer. A 27B decision model that writes no free text, for developers building triage and review flows.

Made by mukaddamzaid on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Amos 1 large is a 27B decision model. You give it a case as text, JSON or a photo, plus a set of typed questions, and it answers every question in one pass. Each answer comes with a probability for every option you allow. Because it returns only answers and no prose, you never parse a reply, and a broken format cannot occur. It is the largest of the four Amos 1 models, meant for the hardest calls.

Who it suits: It suits developers who turn messy cases into structured decisions, such as routing a message or checking a document against a policy.

Industry | Who uses it | How they use it | Example
Customer support | Support leads and developers | Classify and route incoming messages | Ask whether a customer wants a refund and which team should handle it
Back office | Operations teams | Check invoices and records against rules | Read an invoice and answer typed questions about it
Compliance | Analysts | Test cases against long policies | Answer yes or no questions that need several steps of reasoning
Software | Developers | Replace free-text parsing with typed answers | Receive a probability per option instead of a paragraph to parse

If you need free-form writing or a language other than English, you can skip it.

Where to use it: You can use the author's hosted option, which the card says needs no key, or run the serving package yourself on a machine with an 80 GB GPU.

• A hosted API from the author
• Your own server, using the serving package in the repository
• Windows, macOS or Linux with Python, for the download step

Use case | Who | What they get back
Ticket routing | Support teams | A probability for each team and each yes or no question
Document checks | Analysts | Typed answers, quotes or points for each question
Structured extraction | Developers | Output that needs no text parsing

How to set it up: You need Python 3.12, PyTorch 2.14, the Hugging Face hub package and an 80 GB GPU. The card pins those versions to the ones its numbers were measured with.

• Python 3.12 and PyTorch 2.14
• The Hugging Face hub package
• An 80 GB GPU, with CUDA

1. Download the model into a local folder.
2. Install the serving package from that folder.
3. Start the server on port 8000 with the 131072-token limit.

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

How to use it: Send the running server a POST request on its systemone endpoint with a case in state and your typed questions. The card shows this body, which asks two questions about one invoice.

{
  "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}}
  }
}

Tool | What it does
noul | A yes or no question
choice | Pick one from up to 255 options
score | Rate something
multi | Pick any that apply
span | Quote a piece of the text
region | Point at a place in a photo, sent in images

Example: You send an invoice complaint saying a card was charged twice. You ask whether the customer wants a refund and which team should handle it, with billing, technical and sales as options. The reply gives a probability for each answer to each question.

For newcomers and experts: If you are new to this, think of it as a form that fills itself in. You write the questions and the allowed answers, and the model returns how likely each answer is. A first small step is to try the card's invoice example with your own text.

If you know the field, it answers all questions in one forward pass over a 27B Qwen base model, with answer heads in heads.pt and quote decoding settings in calibration.json. The card reports results on its typed-decisions test split and on Decision Bench, and notes that it ran rival models itself, except one whose number is its makers' own.

Case as text, JSON or a photo
Typed questions are added
One forward pass over the model
A probability for every allowed answer

Good to know: The card lists these limits. It is the slowest of the four Amos 1 models, long quotes such as a whole clause can come back cut short, and it handles English only. One request can be as long as 131,072 tokens.

Title: Typed decisions accuracy, as reported on the card
Microsoft-Decision-1: 79.5 percent
Decider 1: 76.8 percent
Amos 1 large: 76.1 percent
Perplexity Decider: 73.9 percent
Clef: 72.5 percent

• License: Apache-2.0
• Cost to run: no API key is named, since the hosted option needs none, but running it yourself needs an 80 GB GPU
• Data it reaches: the hosted option sends your case to the author's server, while the self-run server works on your own machine
• Not for: writing free text or handling languages other than English

On Hugging Face, 13 November 2026

• Formats: Safetensors
• Downloads in the last 30 days: 694
• Likes: 28
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.8-27B
• Tags: safetensors, system-one, decisions, classification, calibration, vision, decision-model

Links

• Model: https://huggingface.co/mukaddamzaid/amos-1-large
• Files: https://huggingface.co/mukaddamzaid/amos-1-large/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mukaddamzaid/amos-1-large

Then follow the usage in its model card. Nothing in this kit runs by itself.