ModelsNews
mammoth-1 and elephant-1: Baseten's open decision models
Baseten's research team, Base Labs, released two open decision models: the 27B mammoth-1 tops the Jev Decision Index, and the 9B elephant-1 leads models its size.
By: Mikola Máté Patrik · · 2 min read
Mikola Máté Patrik
What changed?
On October 21, Baseten released two decision models (Baseten). A decision model does not generate text: it takes a state and a set of questions, then returns a calibrated probability for every option from one shared-prefix prefill. The category emerged in September with TypeSafe's Jev model and System One API.
The numbers
- mammoth-1 (27B): 71.24 on Jev Decision Index 0.4, ranked #1 of 208 models and 1.8 points ahead of Perplexity Decider v1.2.
- elephant-1 (9B): 62.65, the best model under 15B and tied for 8th overall.
- On Base Labs' replica of the Artificial Analysis Decision Index they score 65.9 and 65.2, against 65 for the previous leader, d1-ultra.
- Both are full fine-tunes of Qwen3.5 on 6.21 million decision examples, with a 16,384-token context.
According to the technical report (Base Labs), the team trained the models in one week and explicitly taught them when to answer "I don't know". The full-attention layers use a block-bidirectional mask.
What it is for
Ticket routing, spam and safety filtering, scoring content against a rubric: any task where a full reasoning model would be slow and expensive. Weights are on Hugging Face (mammoth-1, elephant-1), and both models are free to try for a week on Baseten Model APIs with zero data retention by default. If you run many small yes/no decisions through an LLM, benchmark them on your own data. Results come from the team's own runs and the Jev leaderboard.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About
