mammoth-1
mammoth-1A 27B decision model from Base Labs that returns probabilities over predefined options without generating text, built for classification, routing, and scoring.
Model details
View repositorymammoth-1 is a 27B decision model from Base Labs, fine-tuned from Qwen3.5-27B to return probabilities over predefined answers rather than generate text. Given an input and a set of questions, it produces structured decisions for tasks such as classification, routing, and scoring without requiring applications to parse a generated response.
mammoth-1 supports multiple-choice, yes/no, and ordinal scoring questions. Its block-bidirectional attention lets each question incorporate its available options, while shared-prefix caching reuses the input across questions. Per-option probabilities let applications set thresholds for accepting a decision or escalating it for further review.
On release (October 2026), mammoth-1 is the top decision model on the Jev Decision Index.
mammoth-1 leads the Jev Decision Index 0.4 with a score of 71.24. On the Artificial Analysis Decision Index, mammoth-1 scores 65.9.1
2curl --fail-with-body \
3 -X POST "https://inference.baseten.co/v1/systemone" \
4 -H "Authorization: Api-Key $BASETEN_API_KEY" \
5 -H "Content-Type: application/json" \
6 -d '{
7 "model": "decision-models/baseten-mammoth-27b-latest",
8 "state": {
9 "message": "I was charged twice. Please refund the duplicate charge."
10 },
11 "questions": {
12 "route": {
13 "type": "choice",
14 "instructions": "Which team should handle this request?",
15 "criteria": {
16 "billing": "Payments, charges, and refunds",
17 "support": "Technical troubleshooting"
18 }
19 }
20 }
21 }'
