Karpo Decision 31B · open weights
Every answer in one forward pass
A decision model for agents. Give it a state and typed questions; it returns a probability for every option of every question, without generating a token.
Showcases
- 73.86
- Decision Index 0.4.1, Full score: rank #1 (tied)
- 94 ms
- median latency on one GPU, one request at a time
- 4.8
- points of calibration error
- 255
- options a single question can have
From the Decision Index board of 11 November 2026, where the Full score is measured by the maintainers on private tests.
How it works
01
A readout, not generation
Every option is labelled with an option code. A linear readout scores the codes and a softmax over the question's own options gives the probabilities. Nothing is decoded, so the answer is always one of the options.
02
Many questions, one forward pass
A request is one state and any number of questions, encoded as one sequence: the list of questions, the state, then one branch per question with its options.
03
Branch attention
A branch reads the shared block and itself, never another branch. The state is computed once per sequence, and no question sees another question's options.
Three kinds of question
- choice
- One of up to 255 options, each with a key and a description. The answer carries a probability per key.
- noul
- True or false. The answer is the probability of true.
- score
- A level on a scale you describe, lowest first. The answer carries the expected level and a probability per level.
Run it yourself
The weights take 62.5 GB on one GPU. The server answers POST /v1/systemone with the same request the playground shows under “Request”.
pip install "karpo-decision @ git+https://github.com/machinepulse-ai/karpo-decision"
CKPT=$(python -c 'from huggingface_hub import snapshot_download; print(snapshot_download("MachinePulseAI/karpo-decision-31B"))')
torchrun --standalone --nproc_per_node 1 -m karpo_decision.serve.server --checkpoint "$CKPT" --port 18440
