baselabs/elephant-1

A "decision model" post-trained by Base Labs, a subsidiary of Baseten, from Qwen3.5-9B: It takes a state (text, optionally with images) and a set of questions as input, performs a single shared-prefix prefill, and outputs calibrated probabilities for each option of every question. It can be used for ticket routing, classification, and guardrail decisions.

2026-10-28T01:51:05Z

Upvotes 74 · Likes 75 · Downloads 1344

Reasons

Three question types: Yes/No (returns P(yes)), multiple-choice (returns probabilities for each option), and ordinal scoring (returns probabilities for each level); the training objective adds a Brier calibration term alongside cross-entropy, and includes training for an "I don't know" option.
Officially, the Jev Decision Index 0.4 scores 62.65, which is the best among models under 15B and tied for 8th among 208 models; on a replicated version of the Artificial Analysis Decision Index, it scores 65.2, matching d1-ultra (65); the sibling model mammoth-1 at 27B scores higher (71.24).
Training data consists of 6.21M decision samples, of which 16.1% are for safety and abstention; the language component is fully fine-tuned while the visual encoder is frozen. Licensed under MIT, it was not included in previous editions.

Usage

Self-hosting: hf download baselabs/elephant-1 --local-dir elephant-1 && bash elephant-1/setup.sh, which installs a pinned version of the OneJev serving runtime (applying serve.patch); requires one CUDA GPU and the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image.
Invocation: DecisionEngine("baselabs/elephant-1", device="cuda:0").decide(SystemOneRequest(state=text, questions={...})); in questions, specify type (noul / choice / score) and instructions for each question; choice and score types also require criteria.

Risks

Maximum length is 16,384 tokens; requests exceeding this limit are rejected rather than truncated; language tags are English only.
The official Baseten free API week ended on 27 10; one of the two indices is a version replicated by the official team.
It does not use standard Transformers inference; you must run the setup.sh script and patches from the repository.

Open