Tev1-4B-experimental
Together AI's Jev-style classifier. A LoRA fine-tune of Qwen3.5-4B that reads a state, a question and 2 to 24 options and returns one option letter. Served on Together's platform; recipe published.
Tev1 keeps the language-model head and answers by emitting a single option letter, so it is Jev-inspired rather than a non-autoregressive decision head. Together reports 880 of 1,000 on a development set and 300 of 300 on a policy-transfer check; the development set was reused rather than held out, and the authors say the log-probabilities are not calibrated. The weights sit on Hugging Face without a licence field while their licence is finalised, so this registry links to them and does not mirror them. The training code is MIT. A Tev1-0.8B-experimental on Qwen3.5-0.8B followed on 25 September, also without a licence field while its licence is finalised.
What it decides
- choice — picks one option from a set
- 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 | Together AI |
| Released | 2026-09-23 |
| License | Unspecified — weights licence being finalised |
| Reported accuracy | 88.0% |
Hosted API
Served by Together AI — $0.042/MTok input, $0/MTok output. Get access.
Get the weights
pip install systemonemodels
systemone pull together-ai/tev1
The files are served from the maker's Hugging Face repository, togethercomputer/Tev1-4B-experimental, and verified against the checksums recorded here.