Baseten Labs released mammoth-1 and elephant-1 as the world’s best decision models. mammoth-1 ranks first on the Jev Decision Index at 71.24. elephant-1 is the top 9B model on the same index. mammoth-1 scores 65.9 on the Artificial Analysis Decision Index. elephant-1 scores 65.2 there, tying the prior best. Both models use a block-bidirectional attention mask for full attention layers. They were post-trained on Qwen3.5-9B and Qwen3.5-27B. The training data mix contained 6.21M examples across four categories. The models output probability distributions over predefined options for choice, yes/no, or score questions. They avoid text generation and parsing by design. The architecture allows shared prefix caching for efficiency. The data mix was iteratively refined to address failure modes. Both models are now available.

• 71.24 Jev Decision Index score for mammoth-1
• 65.9 Artificial Analysis Decision Index score for mammoth-1
• 6.21M training examples in Mix 8
• 72.2 public Jev Decision Index score for elephant-1