Base Labs has released two open-weight “decision models”, mammoth-1 and elephant-1, and says the larger one ranks first of 208 entries on the Jev Decision Index. The lab published the details on Baseten’s research blog on 21 October 2026, along with downloadable weights for both.
A decision model is not a chatbot. You hand it a record, such as a support ticket, plus one or more questions. It reads everything once and returns a probability for each possible answer. A question can ask for a pick among up to 255 options (“which team handles this?”), a yes or no, or a score on a scale like calm to furious. Each answer comes with a confidence figure.
The pitch is cost and control. A regular language model can make the same call, Baseten’s post says, but it has to write out text that software then parses, and that text says little about how sure the model was. A decision model generates nothing, so it is cheap. Because the odds come back as numbers, a company can set a rule: act automatically above a chosen probability, send a person the case below it. Each extra question about the same record costs only a small additional pass.
Both models are retrained versions of Alibaba’s Qwen3.5, at 9 billion and 27 billion parameters. Base Labs changed how the models read: one attention setting lets each question see its own answer options, which lifted a test score from 44.8 to 48.6 in a single run. Most of the gain, the team says, came from data. Its final training set held 6.21 million examples, and the score of the small model on the public part of the Jev index rose from 45.5 to 74.0 across eight versions of the mix.
Now the scores, and who produced them. On version 0.4 of the Jev Decision Index, mammoth-1 scores 71.24, ahead of Perplexity’s Decider v1.2 at 69.42. Elephant-1 scores 62.65, tied for seventh overall and top among models under 15 billion parameters. According to the post, the board’s maintainers compute that score on private tests, so Base Labs did not grade itself there. The same post admits both models are narrowly beaten in three of five skill areas on those private tests.
The second benchmark is the Artificial Analysis Decision Index. Base Labs says it worked with Artificial Analysis before release, and elephant-1 scores 65, level with Liquid AI’s d1-ultra. The 65.9 for mammoth-1 and 65.2 for elephant-1 shown in the post’s lead chart come from Base Labs’ own replication of that test. Only the 65 for elephant-1 is described as an official result. The mammoth-1 figure on this index is therefore self-reported.
Everything about training is also the team’s own account: the data recipe, the checks that screen out bad examples, and the claim of zero overlap between training data and test sets. The post calls the effort a one-week sprint.
Availability is straightforward. The weights sit on Hugging Face under the baselabs name for both models. The post gives no price, no hosted service, and no license terms, so anyone planning to ship these should read the license on the model page first. Both models read text only; image and video support is listed as future work.
Teams that already run a language model to sort tickets or flag risky requests can test this cheaply: take a few thousand labeled cases, run elephant-1 on them, and see how often its high-confidence picks are right, because the self-reported calibration is the claim the leaderboards do not check.
Reported by Baseten on its research blog on 21 October 2026.
