Baseten's mammoth-1 and elephant-1: open decision models that return probabilities, not text

Base Labs, the research team at Baseten, released two decision models on 21 October: mammoth-1, post-trained from Qwen3.5-27B, and elephant-1, from Qwen3.5-9B. Both are open weights under the MIT license.

What a decision model does

A decision model does not write an answer. You give it a state, such as a support ticket, and one or more questions, and it returns a probability for every option. Base Labs supports three question types:

• yes or no, for example “is this a refund request?”;
• choice, for example which team should own the ticket;
• score, an ordered scale such as “can wait”, “today” or “blocking now”.

All questions are answered from one shared prefill, with no text generation. The point is cost and control: you can route automatically above, say, 90% confidence and send the rest to a person. The models were also trained with an “I don’t know” option, and Base Labs says requests longer than the 16,384-token limit are refused rather than cut short.

This is a growing category. We have covered Cloudflare’s Clef and Liquid AI’s d1.

The numbers, all Base Labs’

Model	Jev Decision Index 0.4
mammoth-1	71.24
Perplexity Decider v1.2	69.42
elephant-1	62.65
Jev 1.14	62.2
Cloudflare Clef	55.7

Base Labs says mammoth-1 ranks first of 208 models on that index. On its own replica of the Artificial Analysis Decision Index it reports 65.9 for mammoth-1 and 65.2 for elephant-1, against d1-ultra’s official 65. Baseten also claims “up to a 10x speed up” over a reasoning model. We have not rerun any of this.

Training, per the report: 6.21 million decision examples, a calibration term in the loss, and text only, with the Qwen vision encoder frozen. Base Labs says it found zero contamination between its training data and the indexes it reports.

How to try it

• Hosted: free for one week on Baseten Model APIs, with zero data retention by default, until 27 October at 11:59 PM Pacific. Vercel’s AI Gateway is also offering both free through 27 October.
• Self-hosted: the model card’s setup.sh installs a pinned serving runtime on one CUDA GPU.

Base Labs’ own advice is the right one: test accuracy on hard cases, check whether its confidence matches how often it is right, and measure cost per decision at your real traffic.