Karpo Decision 31B: open-weight model returns all option probabilities in one 94ms forward pass

ZabihullahAtal · x · 2026-11-18

Karpo released six open-weight decision models (0.8B–31B) that return a probability for every option in a single forward pass—no token generation. Karpo Decision 31B scores 73.9 on Decision Index (tied #1), runs at 94ms median latency on one RTX PRO 6000, and beats Gemini Decision 1 on JevBench (87.2 vs 78.5). The team openly reports overfitting: smallest private same-skill drop among top-10 public scorers (4.3 vs 6.8 avg). Suited to search, recommendations, tool choice, model routing, and browser actions.

Original post →

More from Models

Models channel →