Karpo-Decision-31B drops, tops Decision Index
MachinePulse has open-sourced Karpo-Decision-31B, an Apache 2.0-licensed decision model fine-tuned from Gemma 4 that evaluates states and typed options in a single forward pass without generating text. The 31.3-billion-parameter model ranks #1 on the Decision Index with a 73.9 Full score, outperforming Gemini Decision 1 on benchmarks like When2Call and Mind2Web.
By swapping autoregressive token generation for a linear classification readout over options, Karpo-Decision-31B demonstrates that discrete decision-making and agent routing belong in single forward passes. This architecture offers a much faster, calibrated alternative to using general-purpose LLMs for tool selection and triage.
- –Direct option readout eliminates autoregressive decoding overhead, achieving median latencies around 94 ms per request on a single GPU.
- –Softmax scoring over candidate choices delivers calibrated probabilities with a 4.8-point calibration error for choice, boolean, and rating tasks.
- –Branch attention processes the shared state once while keeping question branches isolated, enabling simultaneous multi-question evaluation.
- –Released under Apache 2.0 with public inference code, the model provides an open-weights foundation capable of rivaling closed models like Gemini Decision 1.
DISCOVERED
6h ago
2026-11-16
PUBLISHED
7h ago
2026-11-16
RELEVANCE
AUTHOR
GithubProjects