mammoth-1

mammoth-1

Answers a set of questions about a situation in one pass and returns a calibrated probability for each option. A 27B decision model from Base Labs at Baseten, tuned from Qwen3.5-27B.

Made by baselabs 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: mammoth-1 is a decision model from Base Labs at Baseten, post-trained from Qwen3.5-27B. You give it a situation as text, optionally with images, plus a set of questions, and it returns a calibrated probability for every option.

All the questions are answered from a single shared prefill of the situation, so adding another question costs little extra work. The probabilities come with a trained option for not knowing.

Who it suits: Teams that want a program to make structured decisions, such as routing or classifying, and to see how confident each answer is.

Industry | Who uses it | How they use it | Example
Customer support | Support platform teams | Route and triage incoming messages | Check whether a customer wants a refund and which team should handle it
Software | Developers of automated workflows | Add a decision step with a confidence score | Pick one of several named options with probabilities
Research | People comparing decision models | Score it on decision benchmarks | Compare it with its smaller sibling elephant-1

Skip it if you need long free-form writing, because it is built to return answers to set questions.

Where to use it:

• Your own machine with one CUDA GPU, using the PyTorch CUDA image the card names
• Docker, through that image
• Baseten Model APIs, as a hosted option
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Ticket routing | Support teams | A team choice with a probability for each team
Yes or no checks | Workflow builders | The probability that the answer is yes
Urgency scoring | Operations teams | A level picked from an ordered list, with probabilities

How to set it up:

• One CUDA GPU
• The pytorch 2.8.0 CUDA 12.8 devel image named on the model card
• Python and the Hugging Face tools for the download

1. Run the download command, which also runs the setup script that installs the pinned serving runtime.
2. Write a short Python script that builds a request with a state and your questions.
3. Run it and read the answers.

hf download baselabs/mammoth-1 --local-dir mammoth-1 && bash mammoth-1/setup.sh

How to use it: Create a DecisionEngine, then call decide with a state and a dictionary of questions. Each question has a type and instructions, and some have criteria.

from qev.engine import DecisionEngine
from qev.schema import SystemOneRequest

engine = DecisionEngine("baselabs/mammoth-1", device="cuda:0")

questions = {
    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {"billing": "Charges, refunds, invoices",
                     "technical": "App or site faults",
                     "fraud": "Suspected unauthorised use"},
    },
    "urgency": {
        "type": "score",
        "instructions": "How urgent is this?",
        "criteria": ["Can wait", "Today", "Blocking the customer now"],
    },
}
resp, _ = engine.decide(SystemOneRequest(
    state="I was charged twice this month, please refund one of them.", questions=questions))
print(resp.answers)

Question type | Criteria | Answer fields
noul, a yes or no question | Optional | noul, the probability of yes
choice, one of named options | A name and description for each option | choice and probabilities
score, ordered levels | A list of levels, lowest first | score and probabilities

Example: You pass in a message from a customer who was charged twice and ask three questions: is it a refund request, which team should handle it, and how urgent is it. The answers come back with probabilities for each option.

For newcomers and experts: If you are new to this, think of it as a form that fills itself in. You describe the situation, list your questions, and get an answer and a confidence for each. Start with the example script above.

If you know the field, the language model was fully fine-tuned on text and image decisions with the vision encoder frozen, and full-attention layers use a block-bidirectional mask. Training combined cross-entropy with a Brier calibration term, and the maximum sequence length is 16,384 tokens.

Describe the state in text or images
List your questions and their types
One shared prefill of the state
Probabilities for every option

Good to know:

• The authors report first place out of 208 models on their decision benchmark, and these figures come from them
• For quicker and cheaper decisions the card points to the 9B sibling elephant-1
• Weights: about 54.7 GB in BF16, according to Hugging Face

Title: Jev Decision Index 0.4 scores
mammoth-1: 71.24 points
Perplexity Decider v1.2: 69.42 points
elephant-1: 62.65 points
Jev 1.14: 62.2 points
Cloudflare Clef: 55.7 points

• License: MIT
• Cost to run: a CUDA GPU for self-hosting, or the hosted Baseten Model APIs option, which the card says was free for a limited time
• Data it reaches: the state text and images you send, and the hosted endpoint has zero data retention on by default
• Not for: requests longer than 16,384 tokens, which are refused rather than cut

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,903
• Likes: 97
• License: mit
• Languages: en
• Based on: Qwen/Qwen3.5-27B
• Quantized versions: 6
• Fine-tunes: 24
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, decision-model, classification, calibration, routing, guardrails, vision, multimodal, baseten

Links

• Model: https://huggingface.co/baselabs/mammoth-1
• Files: https://huggingface.co/baselabs/mammoth-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download baselabs/mammoth-1

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