Baseten releases Mammoth-1 and Elephant-1 decision models
Baseten has released two post-trained Qwen3.5 models, at 27B and 9B parameters, that it says lead decision-model benchmarks, where models return calibrated probabilities over fixed options instead of generating text.
Baseten published a technical report on two "decision models": mammoth-1 (27B) and elephant-1 (9B). Both are post-trained versions of Qwen3.5-27B and Qwen3.5-9B.
TypeSafe introduced the decision-model format in September 2026 with Jev and its System One API. A decision model takes a JSON state, such as a support ticket or document, and a set of questions. For each question it returns a probability distribution over the options rather than generated text. Questions can be multiple choice (1–255 options), yes/no, or scores on a 2–10 level scale. Because nothing is decoded, Baseten says the outputs need no parsing, cost less to run, and can be thresholded or calibrated for uncertainty.
According to Baseten, mammoth-1 ranks #1 of 208 models on the Jev Decision Index 0.4. The company says elephant-1 is the best 9B model on that index, ahead of Jev 1.14, and ties the previous leader, d1-ultra, on the Artificial Analysis Decision Index.
The report credits two techniques. The first is a block-bidirectional attention mask, applied only to the 16 full-attention layers of Qwen's hybrid Gated DeltaNet architecture. It lets each question attend to its options while the shared prefix cache is reused across questions. In a single run on fixed data, the mask raised the index score from 44.8 to 48.6 on a public 10k sample. The second is an iterative synthetic-data recipe that targets failure modes found in evaluation.
The available excerpt does not include pricing, availability, or licensing details.