Rohan Paul (@rohanpaul_ai)T2·78 pts
- The 31B model ties for first at 73.9 on the 300-entry Decision Index, built on Gemma 4 31B
- One forward pass returns a probability per option with zero generated tokens and always-valid answers
- Beats Gemini Decision 1 on Amazon ESCI (64.7% vs 58.7%) and New Yorker Caption (90.5% vs 79.7%)
- 94 ms median latency on a single RTX PRO 6000, roughly one-fifth lower cost per 1,000 decisions