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.
Featherless AI's open server that turns a stock Hugging Face language model into a typed-decision endpoint. It reads the next-token logits for each question and builds the Choice, Score or Noul answer itself; no classifier head is trained and no JSON is generated.
Decides
choice, score, noul, classify
choice, score, noul, classify, route
Architecture
jevembed
simple-jev
Fine-tuned from
qwen/qwen3-embedding-4b
google/gemma-4-26b-a4b-it
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights + hosted API
Hosted by
—
Featherless AI
Input price
—
—
Decision accuracy
85.9%
—
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 simple-jev?
jevembed is from HIT-TMG (Lychee Team) and simple-jev from Featherless AI. jevembed has open weights you can download and run; simple-jev has open weights and a hosted API. Both answer choice, score, noul and classify questions. Only simple-jev answers route.
Which is more accurate, jevembed or simple-jev?
Only jevembed publishes an accuracy figure (85.9% on JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging); simple-jev does not, so there is no comparison to make without your own test.
Which is cheaper, jevembed or simple-jev?
jevembed: Free (open weights). simple-jev: Hosted, price not published, or free to self-host. Open weights cost nothing per call beyond your own hardware.
Can I run jevembed or simple-jev locally?
Yes, both: systemone pull hit-tmg/jevembed and systemone pull featherless-ai/simple-jev download the weights.
—
Evaluation suite
JevEmbed-Data test split (64,110 hard-label questions), final LoRA checkpoint before merging
