An embedding model turned decision model. Qwen3-Embedding-4B with a merged LoRA, driven by the open JevEmbed toolkit, which scores Choice, Score and Noul questions from embedding similarity and returns a distribution without generating text.
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, score, noul, classify
choice, score, noul, classify, route
Architecture
jevembed
julia
Fine-tuned from
qwen/qwen3-embedding-4b
jhu-clsp/mmbert-small
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
85.9%
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 jevembed and julia-1?
jevembed is from HIT-TMG (Lychee Team) and julia-1 from Supersonic Labs. Both have open weights you can download and run. Both answer choice, score, noul and classify questions. Only julia-1 answers route. julia-1 reads up to 8K tokens of state, against 1K tokens for jevembed. julia-1 is the smaller model, at 144M parameters to 4.0B.
Which is more accurate, jevembed or julia-1?
They report on different suites — jevembed 85.9% on JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging, 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, jevembed or julia-1?
jevembed: Free (open weights). julia-1: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run jevembed or julia-1 locally?
Yes, both: systemone pull hit-tmg/jevembed and systemone pull supersonic-labs/julia-1 download the weights.
Evaluation suite
JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging
typed-decisions test set (400 cases, 2,000 questions)
