Tev1-4B-experimental

Tev1-4B-experimental is a small decision model from Together AI, published on 23 September 2026 with a blog post titled “How to train your own Jev for $17”, a Hugging Face repository and the training code. It is Together’s open reproduction of the “System One” idea behind TypeSafe’s Jev: given a structured description of a situation, a question and between two and 24 possible answers, it returns the letter of one answer rather than writing text. It is a low-rank fine-tune of Qwen3.5-4B, about 4 billion parameters, trained on sequences of up to 2,048 tokens; Together has not stated a context limit for use. The point of the release is the recipe as much as the model: Together shows the whole training run costing about $17, which tells a buyer that a decision model tuned on their own labelled decisions is within reach. Together’s development results — 880 of 1,000 main decisions and 300 of 300 policy-transfer decisions — are described on its own model card as development results, not an independent benchmark, and are not stored as scores. The code and documentation are MIT-licensed; the licence for the weights is described as still being finalised, so commercial use of the weights should wait for it.

Not yet independently verified: Together AI’s blog, model card and repository are the only sources for this model, apart from one independent tracker; no press report was found. The price of $0.042 per million input tokens appears only in Together’s post on X, not on its pricing page, so it is not recorded. The licence for the weights is described as still being finalised. We will update this when it can be confirmed, and remove this note.