Tev1-4B-experimental

Most language models write text. Tev1-4B-experimental, from Together AI, instead returns the letter of one answer from a list. It was published on 23 September 2026 alongside a blog post, a Hugging Face repository and the training code.

A small decision model

Together presents it as its open take on TypeSafe's Jev, built on the System One approach. The model receives a structured account of a situation and a question, along with candidate answers numbering from two up to 24, and chooses one. Technically it is a low-rank fine-tune of Qwen3.5-4B, with around 4 billion parameters, trained on sequences no longer than 2,048 tokens. Together gives no context limit for use, and it labels the release experimental.

The recipe is the point

The release matters as much for its recipe as for the model. Together's post walks through an entire training run for a very small outlay, suggesting that a buyer could tune a decision model on their own labelled decisions. Because the training code is published too, others can reproduce the run on their own data. In its development results Together reports getting 880 of 1,000 main decisions and all 300 policy-transfer decisions right. Its own model card describes these as development results, not an independent benchmark.

Licensing and sources

Code and documentation carry the MIT licence, while the licence for the weights is still described as being finalised, which argues for holding off on commercial use of the weights. Together's blog, model card and repository are the only sources for the model apart from one independent tracker, and no press report was found. A price per million input tokens appears only in a social-media post rather than on Together's pricing page, so it is not a dependable figure.