Tev1-4B-experimental

Tev1-4B-experimental is a small open decision model from Together AI, published on 23 September 2026 with its training code. It is Together's open version of the System One idea behind TypeSafe's Jev. Presented with a structured description, a question and anywhere from two to 24 options, it outputs the letter of the chosen answer instead of writing text.

What Tev1-4B-experimental offers

• Choosing an option: returns the letter of one answer from the list supplied.
• Training recipe: Together publishes the full recipe, so a team can train a similar model on its own labelled decisions.
• Base model: built by lightly adapting Qwen3.5-4B, a model of roughly 4 billion parameters, with training on sequences up to 2,048 tokens.
• Development results: 880 correct out of 1,000 main decisions, and 300 out of 300 policy-transfer decisions, on the model card rather than an independent benchmark.

Key facts

• Developer: Together AI.
• Code licence: MIT for the code and documentation.
• Weights licence: still being finalised, and commercial use should wait for it.
• Hosting: the model is published on Hugging Face.