Biplov GautamTogether AI released Tev1-4B-experimental on 23 September 2026 with something the other System One launches did not include: a tutorial showing how it was trained, and a bill. The training run cost about $17. That, more than the model, is the point Together is making — a decision model in Jev's shape is a weekend project on a serverless GPU.
How it answers
Tev1 keeps Qwen3.5-4B's language-model head. Given a state, a question and between 2 and 24 lettered options, it emits a single option letter. So it is Jev-inspired rather than non-autoregressive: one token is generated, and the probability over the letters at that position is the distribution. Together says plainly that those log-probabilities are not calibrated.
The fine-tune is a LoRA over roughly 38,000 training examples (the GitHub README says 37,840, the blog 38,340). The weights on Hugging Face include the base's vision tower, which is why the safetensors total 4.66B parameters.
The numbers
Together reports 880 of 1,000 on its development set and 300 of 300 on a policy-transfer check. The development set was reused rather than held out, so the 88% is a fit, not a generalisation number, and no independent benchmark has been published yet. The registry shows it as reported, with that caveat.
Getting it
Tev1 is on Together's platform — the model page cites $0.042 per million input tokens with free output, the same price as Jev — and the weights are on Hugging Face. But the Hub entry has no licence field: the README says the weights licence is being finalised before public conversion. Until it is, the registry links to the files and does not mirror them.
systemone pull together-ai/tev1 # fetches from Hugging Face; read the licence first
The training code is MIT, so the recipe is yours to reuse even where the weights are not.
Why it matters
Tev1 shows the floor of the field: a competent Jev-style model from an open base, a small dataset and $17 of compute. If you need calibrated probabilities and open weights today, Kev and Nimble start from the same Qwen family and publish their calibration. Tev1's page will carry the licence the moment Together sets one.