Jevling-E2B-v1
A small System One model on Gemma 4 E2B: a text state and several typed questions (choice, yes/no, score) go in, and one forward pass returns a probability per option, with no generated text. Tuned for on-device kiosk ordering in Traditional Chinese. Apache-2.0.
Jevling reads each answer from a letter slot after the question, so all of a state's questions are answered in one pass; the fine-tune trains only that slot and the base can still chat. Calibration is a temperature (1.10) folded into the final norm, and no calibration error is reported. On JevBench's 231 public items (a third-party board) the maker's own run gives 100% easy, 94.4% original and 43.2% hard; the maker also reports 15 public datasets and a zh-TW kiosk set that is its own synthetic data. The maker says v1 was retrained only on commercially usable data. Sibling Jevling-0.8B-v1 on Qwen3.5-0.8B. Both have Q8_0 GGUF builds that need BRICKS's llama.cpp fork (mybigday/system-one-llama.cpp); the maker warns that stock llama.cpp loads them but answers wrongly without an error.
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 probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 5.1B |
| Base model | google/gemma-4-e2b-it |
| Maker | BRICKS |
| Released | 2026-10-03 |
| License | apache-2.0 |
Get the weights
pip install systemonemodels
systemone pull bricks/jevling
The files are served from the maker's Hugging Face repository, BricksDisplay/jevling-e2b-v1, and verified against the checksums recorded here.