mev
Mellow AI's Jev-inspired decision model: a fine-tune of Qwen3.5-4B on Together's open Tev1 recipe that reads a state, a question and 2 to 24 options and returns one option letter. Option probabilities come from the first generated token.
It keeps the language-model head and emits a single option letter, so it is Jev-inspired rather than a non-autoregressive decision head, the same approach as Together's Tev1; the card says it does not reproduce Jev's runtime and was not trained on Jev outputs. A rank-8 LoRA on all linear modules, one epoch on the 37,840-example Tev1 recipe mixture, published merged. The maker's numbers are on development sets reused from that recipe, not held out: 86.7% on the main decisions set against Tev1's 88.0% on the same set, with ECE 0.038. Yes/no maps to two options; score is computed client-side. English only; no hosted API. Training-data sources keep their own terms.
What it decides
- choice — picks one option from a set
- 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 | 4B |
| Base model | qwen/qwen3.5-4b |
| Maker | Mellow AI |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 86.7% |
Get the weights
pip install systemonemodels
systemone pull mellow-ai/mev
The files are served from the maker's Hugging Face repository, stay-mellow-ai/mev, and verified against the checksums recorded here.
Read more
This page was opened by System One for Mellow AI, who can claim the organisation and take it over at any time.