Baseten releases elephant-1, a 9B decision model matching d1-ultra at a quarter of the cost

Details

Base Labs at Baseten released elephant-1, a 9B decision model post-trained from Qwen3.5-9B that returns probabilities for yes/no, multiple-choice, or ordered-score questions rather than generating text.
Artificial Analysis scored the model 65 on the Decision Index, tying d1-ultra, with a cost of $8 per 1M tasks compared to roughly $31 for d1-ultra.
The model excels at routing and guardrails with 95 on When2Call and 93 on WildGuardTest, but has only 12 accuracy on Omniscience due to a high rate of "I don't know" responses.
Architectural changes include swapping causal masks for block-bidirectional ones on full-attention layers to allow question blocks to see all options and reuse prefix caches.
Training used 6.21M examples with a mix of 58% core decision skills, 16% safety, 16% judging, and 10% public datasets, optimized with cross-entropy and a Brier loss term.
Weights are available as open weights on Hugging Face, with a 16k context limit and free API access on Baseten until October 27.

Key point