Baseten's research lab releases Mammoth-1 and Elephant-1 decision models

Baseten's internal research team, Base Labs, released two new "decision models" called mammoth-1 (27 billion parameters) and elephant-1 (9 billion parameters). Unlike typical chat models that write out text, decision models return a probability score across a fixed set of possible answers, so they can judge choices, ratings, or yes/no questions directly without generating a written response.

Both models are fine-tuned versions of Qwen3.5, trained on 6.21 million decision-making examples combining public and synthetic data. Baseten reports mammoth-1 ranks first out of 208 models on the Jev Decision Index with a score of 71.24, and that elephant-1 is the top-ranked 9B model on that index and tied for first on a separate index called the Artificial Analysis Decision Index. These are the company's own benchmark claims.

Baseten pitches the models for tasks like support-ticket routing, spam and phishing detection, and jailbreak or guardrail checks, since skipping full text generation can make the models cheaper and faster than using a general chat model for simple classification, and the company says confidence thresholds can control when a decision is automated versus escalated to a person. Both models are open-source on Hugging Face and free to try for a week through Baseten's Model APIs.

Why it matters

It offers developers a cheaper, faster alternative to full LLMs for routine classification and routing tasks baked into production systems.

Source: Baseten