A 0.6B decision model on Qwen3-0.6B-Base from the DocsGPT team. Give it a state and typed questions (yes/no, a choice of up to 16 options, a 3- or 4-level score) and it returns calibrated probabilities in one pass, for RAG and agent checks.
A 0.4B decision model made from the first 20 layers of Qwen3-0.6B plus a small attention head that compares options. It answers choice, noul and score questions in one pass, and choice order cannot change its answer by construction. Non-commercial licence.
Decides
choice, score, noul, classify, rank, route
choice, score, noul
Architecture
arc-decide
bev-decider
Fine-tuned from
qwen/qwen3-0.6b-base
qwen/qwen3-0.6b
License
mit
cc-by-nc-4.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
—
74.7%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
—
p95 latency
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between arc-decide and bev-decider?
arc-decide is from Arc53 and bev-decider from Avishek Biswas. Both have open weights you can download and run. Both answer choice, score and noul questions. Only arc-decide answers classify, rank and route. bev-decider is the smaller model, at 478M parameters to 600M. arc-decide is licensed mit; bev-decider, cc-by-nc-4.0.
Which is more accurate, arc-decide or bev-decider?
Only bev-decider publishes an accuracy figure (74.7% on avbiswas/bev-decision test split (5,000 held-out questions over 2,617 states), the maker's own); arc-decide does not, so there is no comparison to make without your own test.
Which is cheaper, arc-decide or bev-decider?
arc-decide: Free (open weights). bev-decider: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run arc-decide or bev-decider locally?
Yes, both: systemone pull arc53/arc-decide and systemone pull avishek-biswas/bev-decider download the weights.
—
Evaluation suite
—
avbiswas/bev-decision test split (5,000 held-out questions over 2,617 states), the maker's own
