JevEmbed-Qwen3-Embedding-4B
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.
Trained for one epoch on the 1.6M-question JevEmbed-Data training split (LoRA rank 64, 16 GPUs). On that dataset's 66,482-question test split the final checkpoint reaches 85.9% overall hard-label accuracy (Choice 90.3%, Score level 73.4%, Noul 95.9%), against 36.3% for the untuned base. Inputs are truncated at 1,024 tokens. No JevBench result is published. Siblings on the same recipe: JevEmbed-Qwen3-Embedding-0.6B and JevEmbed-KaLM-Embedding-V2.5. The toolkit (Apache-2.0) also covers data synthesis, fine-tuning, a benchmark module and a playground. Training-data source licences vary; see the dataset's processing report.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3-embedding-4b |
| Maker | HIT-TMG (Lychee Team) |
| Released | 2026-09-28 |
| License | apache-2.0 |
| Reported accuracy | 85.9% |
Get the weights
pip install systemonemodels
systemone pull hit-tmg/jevembed
The files are served from the maker's Hugging Face repository, HIT-TMG/JevEmbed-Qwen3-Embedding-4B, and verified against the checksums recorded here.