mammoth-1

// Overview

mammoth-1 is a 27B-parameter decision model from Base Labs, the research lab at Baseten, announced on Baseten's blog on October 21, 2026 together with its 9B sibling elephant-1. It is post-trained from Qwen3.5-27B. A decision model does not write text: it takes a state (text, optionally with images) and a set of typed questions — noul for yes/no, choice for one of several named options, score for ordered levels — and returns a calibrated probability for every option, answering all the questions from one shared-prefix prefill. The weights are on Hugging Face under the MIT license, and the model is served on Baseten Model APIs.

On the Jev Decision Index 0.4, mammoth-1 scores 71.24, which its model card ranks first of 208 models, ahead of Perplexity Decider v1.2 (69.42), elephant-1 (62.65), Jev 1.14 (62.2) and Cloudflare Clef (55.7). On the public split it leads in four of the five areas: knowledge and reasoning, language, tools and automation, retrieval and classification, and arts and human taste. On Base Labs' replica of the Artificial Analysis Decision Index — a weighted average of AA-Omniscience-Decision, HLE-MultiChoice, When2Call and WildGuardTest — it scores 65.9, against 65.2 for elephant-1 and 65 for Liquid AI's d1-ultra.

The technical report describes the recipe. Both models are a full fine-tune of the language model on text and image decisions with the vision encoder frozen, and the full-attention layers switch from a causal to a block-bidirectional mask (the Gated DeltaNet layers stay causal); in a 9B ablation that change lifted the index from 44.8 to 48.6. Training used 6.21M decision examples ("Mix 8": 58.3% core decision skills, 16.1% safety and abstention, 16.1% judging and expert domains, 9.5% public decision data), a cross-entropy loss plus a Brier calibration term, and a trained "I don't know" option. mammoth-1 ran one epoch of 9,859 steps on 18 nodes of 4 GB300 GPUs. The maximum sequence length is 16,384 tokens, and longer requests are refused rather than truncated.

On Baseten Model APIs the model id is decision-models/baseten-mammoth-27b-latest, called through the /v1/systemone endpoint at inference.baseten.co, with zero data retention on by default; the model card and blog offered it free for one week after launch, through October 27, 2026. To self-host, the Hugging Face repository ships setup.sh, which installs a pinned serving runtime (OneJev plus a patch) on one CUDA GPU, and a Python DecisionEngine for local calls.

// Benchmarks

This model's scores

1. Jev Decision Index 0.471.24
2. Artificial Analysis Decision Index (Base Labs replica)65.9

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• 71.24 on the Jev Decision Index 0.4, ranked first of 208 models on its model card
• Leads four of the five public Decision Index areas, from knowledge and reasoning to arts and human taste
• 65.9 on Base Labs' replica of the Artificial Analysis Decision Index, ahead of d1-ultra at 65
• Text and image states, with a probability for every option instead of generated text
• Every question about a state answered from one shared-prefix prefill
• MIT-licensed weights on Hugging Face, plus a hosted API with zero data retention by default

// Best for

• Reach for mammoth-1 to route, triage or classify requests: one call answers a yes/no, a choice and a score question about the same ticket, with probabilities you can threshold.
• Reach for it as a guardrail or LLM-as-a-judge in an agent, where you act on calibrated probabilities and a trained "I don't know" option rather than parse text.
• Reach for it for decisions about images as well as text, such as visual checks on a screenshot or a photo.
• Pick elephant-1, its 9B sibling, when latency and cost matter more than the last points of accuracy.
• Look elsewhere when you need generated text, or for states longer than 16,384 tokens, which the model refuses.

// How to access

// mammoth / elephant (decision models) — every version

The full lineage of the mammoth / elephant (decision models) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.