JevAlt Deem-4B
Three 4B decision models by Mert Kaya, fine-tuned from Intern-Decision-4B for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B). They speak the Jev API with choice, score and noul questions, can answer "unknown", and the Q4 GGUF runs on a CPU in about 3 GB.
A personal project published under the Eschatia Labs label. A rank-32 LoRA merged into Intern-Decision-4B per language, with a fitted calibration temperature and optional conformal sets; no ECE is published and the card says to recalibrate on your own data. Optional reasoning (off by default) writes up to 256 tokens before answering. No image input, although the vision tower is present. All numbers are the maker's own, on held-out splits of his own teacher-labelled data pipeline, which the card says favour JevAlt. Not affiliated with TypeSafe; unrelated to LibertAI's Deem.
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
- route — sends the input to one of several destinations
At a glance
| Parameters | 4.5B |
| Base model | internlm/Intern-Decision-4B |
| Maker | Mert Kaya |
| Released | 2026-10-01 |
| License | apache-2.0 |
| Reported accuracy | 94.7% |
Get the weights
pip install systemonemodels
systemone pull mert-kaya/jevalt
The files are served from the maker's Hugging Face repository, mertkayacs/Deem-4B, and verified against the checksums recorded here.