Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer

CodeByPoonam · x · 2026-11-16

Karpo released a family of six open-weight decision models (0.8B-31B) for agents: given a state and typed questions, it returns probabilities for every option in a single forward pass without generating tokens. Karpo-Decision-31B scores 73.9 on the Decision Index (tied #1), 90.4% on When2Call vs Gemini Decision 1's 79.9%, and 87.2 on JevBench Capability vs 78.5, with 94ms median latency on one RTX PRO 6000 (62GB weights). Key design: a linear readout with softmax over option codes; branch attention keeps questions independent in one sequence; supports choice (up to 255 options), bool, and score question types. Suggested uses: search, recommendations, tool choice, model routing, browser actions.

Original post →

More from coding & agent

coding & agent channel →