Ruoxi Qiu's decision adapter for a frozen Qwen3.8-27B: a 6.3M-parameter LoRA on the last four text layers plus a numerical prior. One shared state feeds isolated question branches; each Choice, Noul or Score question returns probabilities with no generated tokens.
A 144M-parameter System One model on the multilingual mmBERT-small encoder. Takes a state, a question and 2 to 20 options and returns one decision with probabilities for choice, score and yes/no questions, on a CPU.
Decides
choice, noul, score
choice, score, noul, classify, route
Architecture
canopy-jev
julia
Fine-tuned from
qwen/qwen3.8-27b
jhu-clsp/mmbert-small
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
87.4%
73.2%
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 canopy-jev and julia-1?
canopy-jev is from Ruoxi Qiu and julia-1 from Supersonic Labs. Both have open weights you can download and run. Both answer choice, noul and score questions. Only julia-1 answers classify and route. julia-1 is the smaller model, at 144M parameters to 27B.
Which is more accurate, canopy-jev or julia-1?
They report on different suites — canopy-jev 87.4% on JevBench public set (231 items; 202 correct), the maker's own fresh zero-shot run with the upstream scorer, julia-1 73.2% on typed-decisions test set (400 cases, 2,000 questions) — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, canopy-jev or julia-1?
canopy-jev: Free (open weights). julia-1: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run canopy-jev or julia-1 locally?
Yes, both: systemone pull camellia86/canopy-jev and systemone pull supersonic-labs/julia-1 download the weights.
Evaluation suite
JevBench public set (231 items; 202 correct), the maker's own fresh zero-shot run with the upstream scorer
typed-decisions test set (400 cases, 2,000 questions)
