Baseten decision models were released on October 21, 2026 by Base Labs, Baseten's research lab: mammoth-1 (27B) and elephant-1 (9B) generate no text at all. Instead, each returns a probability for every candidate answer, with all scores summing to 1. Both models are open source on Hugging Face, free to try for a week through Baseten Model APIs, with zero data retention by default.
They answer three narrow kinds of questions
Decision models handle only three task types: choice, picking one option from a list, such as whether a ticket belongs to infrastructure, billing, security or support; scoring, rating input on a scale, such as whether an email is high or low urgency; and boolean judgments, such as whether an email is a phishing attempt. Simple as they sound, these are among the highest-volume requests in production systems. Teams have traditionally used generative LLMs for them, which must write out an answer word by word: costly, slow, and with no reliable signal of how certain the model actually is.
Confidence is the real product
mammoth-1 ranks first among 208 decision models on the Jev Decision Index 0.4 with a score of 71.24, while elephant-1 leads the 9B class. Both were post-trained from Qwen3.5 on 6.21 million decision examples spanning knowledge and reasoning, language, tools and automation, retrieval and classification, and human taste, including examples that teach the model to say it does not know. Baseten says decision models can be up to 10 times faster than reasoning models on such tasks.
Who should take a close look
The point is not to replace chat models but to sit in front of them: route automatically when confidence is high, say 90% for ticket assignment, and escalate to a person or a stronger model when it is not. Ticket triage, spam and phishing detection, prompt guardrails and jailbreak checks, and routing between small and frontier models are ready-made uses. The trade-offs are real: these models judge without explaining, so audit-heavy workflows still need separate records, and the headline scores come from vendor-reported benchmarks. Test accuracy and confidence calibration on your own hard cases before setting thresholds.
