saved
Mammoth-1 and Elephant-1: frontier decision models
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Baseten Labs releases mammoth-1 and elephant-1, open Qwen3.5 decision models with block-bidirectional attention. mammoth-1 leads the Jev Decision Index 0.4 at 71.24 (of 208 models), and elephant-1 scores 62.65 as the best under-15B entry; on Artificial Analysis, elephant-1 ties d1-ultra at 65.
ideas
- Decision models return calibrated choice distributions in one forward pass. Inputs are a state plus a question with 1–255 options; outputs are per-option probabilities (summing to 1) and a confidence score, aimed at routing and classification without a full generative judge.
- Architecture keeps GDN causal and opens attention within blocks. Built on Qwen3.5-27B (mammoth-1) and Qwen3.5-9B (elephant-1), the team replaces the causal mask only on full-attention layers with block-bidirectional masking so tokens see their whole block and prior blocks.
- Data mix iteration, not just scale, moved the score. Eight mixes totaling 6.21M examples lifted the 9B public Jev board from 45.5 to 66.0, with elephant-1 (Mix 8) at 72.2 on that public suite after decontamination against both indexes.
- Leaderboard results are study-scoped. mammoth-1’s 71.24 beats Perplexity Decider v1.2 (69.42) by 1.8 on Jev 0.4 private tests; AA scores for elephant-1 (65) match Liquid d1-ultra on AA’s 20 October 2026 public leaderboard.
quotes
“We provide the technical report for training the world's best decision models.”
“mammoth-1 leads the Jev Decision Index at 71.24; elephant-1 is the top 9B model.”
“Growing this mixture saw our score improve from 45.5 to 66.0 on this benchmark.”