Baseten open-sources decision models mammoth-1 (27B) and elephant-1 (9B): per-option probabilities instead of text, #1 on the Jev Decision Index

Background

Ticket triage, spam detection, prompt guardrails and model routing only need a pick from a fixed set of answers, yet many teams still have a generative LLM write text and then parse it, which is slow, costly and gives no reliable confidence. Decision models are a new category that skip text generation and return a probability for every option. On Oct 21, Baseten released two open decision models trained by its research team Base Labs: mammoth-1 (27B) and elephant-1 (9B), in the same space as Perplexity's pplx-decider line.

Architecture

Both are full fine-tunes of Qwen3.5-27B and Qwen3.5-9B on 6.21M decision examples mixing public datasets and synthetic data, covering core decision skills, safety and abstention, and judging and expert domains. They support choice, boolean and ordinal score questions. Block-bidirectional attention lets each question attend to its options, and shared-prefix caching reuses one input across many questions, so a single forward pass returns every distribution. Baseten says this can be up to about 10x faster than a reasoning model (vendor claim).

Benchmarks (vendor-reported)

mammoth-1 scores 71.24 on the Jev Decision Index 0.4, #1 of 208 decision models, and 65.9 on the Artificial Analysis Decision Index. elephant-1 scores 62.6 on Jev (tied #8 overall, best 9B) and 65.2 on Artificial Analysis, tied #1 with the previous leader d1-ultra. No independent third-party results have been published yet.

Getting started

Weights are on Hugging Face (baselabs/mammoth-1, baselabs/elephant-1), MIT per the model pages. Baseten Model APIs are free for a week until Oct 27, 11:59 PM PT, with zero data retention by default. Call /v1/systemone with a state and typed questions to get per-option probabilities. Evaluate accuracy and calibration on your own hard cases, then set thresholds for auto-accept versus escalation; pick mammoth-1 for broader knowledge and elephant-1 for speed.