The Karpo Decision family
Score vs. size on the Decision Index 0.3 public suite, board of 12 November 2026.
The family's largest model, karpo-decision-31B, is ranked #1 on the
Decision Index, with the highest Full score, 73.9, on
the maintainers' private tests (edition 0.4.1, board of 12 November 2026, 300 entries); its public index is 75.0.
How it works
- A readout, not generation. Every option is labelled with an option code. At the readout point that ends each
question, a linear readout scores 255 option 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.
- 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 each question again with its options and a readout point.
- Hybrid attention. The full-attention layers read the whole sequence in both directions; the linear-attention
layers read it in order, so the questions come before the state.
The question-list-first layout follows FlexDM and
jev-brain.
Use
Inference code: github.com/machinepulse-ai/karpo-decision, with the qwen extra. Python 3.12 or later and one GPU; the weights
take 16.8 GB, long requests more. Loading the checkpoint as a plain chat model does not reproduce its answers.
pip install "karpo-decision[qwen] @ git+https://github.com/machinepulse-ai/karpo-decision"
from huggingface_hub import snapshot_downloadfrom karpo_decision.decider import Decider path = snapshot_download("MachinePulseAI/karpo-decision-9B")decider = Decider.from_pretrained(path, device="cuda")response = decider.decide({ "state": "Order 1182: the customer asks to return a blender bought 45 days ago. Policy: returns within 30 days.", "questions": { "eligible": {"type": "noul", "instructions": "Is the return eligible under the policy?"}, "action": {"type": "choice", "instructions": "What should the agent do?", "criteria": {"refund": "Issue a full refund", "decline": "Politely decline the return", "escalate": "Escalate to a supervisor"}}, },})print(response["answers"])# {"eligible": {"type": "noul", "noul": P(true)},# "action": {"type": "choice", "choice": <key>, "probabilities": {<key>: p, ...}, "confidence": c}}
A screenshot goes in the state at viewport_screenshot, as {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}.
HTTP server (POST /v1/systemone, GET /healthz):
CKPT=$(python -c 'from huggingface_hub import snapshot_download; print(snapshot_download("MachinePulseAI/karpo-decision-9B"))')torchrun --standalone --nproc_per_node 1 -m karpo_decision.serve.server --checkpoint "$CKPT" --port 18440
Decision Index kit, installed separately:
pip install "decision-index @ git+https://github.com/apolinario/decision-index"python -m decision_index run --engine karpo_decision.serve.engine:DecisionIndexEngine \ --option checkpoint="$CKPT" --out runs/karpo-decision-9B
Lineage
Qwen/Qwen3.5-9B-Base → Qwen/Qwen3.5-9B
→ karpo-decision-9B (fine-tuned)
Limitations
- Images: a screenshot is read through the base model's vision tower; image inputs were not evaluated.
- Not a chat or text-generation model; generation was not evaluated.
- A question that does not fit in 32,768 tokens with its state is refused, never truncated.
- It always picks one of the given options.
Licence
The weights and model files are licensed under the Apache License, Version 2.0.
Copyright 2026 MachinePulse Pte. Ltd.
karpo-decision-9B is a modified version of Qwen/Qwen3.5-9B (Apache-2.0):
decoder weights fine-tuned, readout added. Qwen3.5 stays under its own licence, a copy of which is in licenses/Qwen3.5-LICENSE; see NOTICE.
The inference code (github.com/machinepulse-ai/karpo-decision) is licensed under Apache-2.0.
Attribution
Not required by the licence, but appreciated: if a product or service uses this model or a derivative, show
Powered by karpo-decision (MachinePulse)
with a link to https://huggingface.co/MachinePulseAI/karpo-decision-9B where the medium allows one.