How to train your own Jev for $17

Together AI has launched a classifier called together/Tev1-4B-experimental, built on top of Qwen3.5 4B on its serverless platform. The company described the release in a blog post that also walks through fine-tuning a version of the model.

The post describes Jev as a classification model that is fast and cheap to run, and says it has become one of the most talked about model releases in the AI space. According to Together AI, such a model takes a piece of state plus predefined questions and returns a result as a score, boolean value, or multiple-choice answer. Real-world applications cited include e-commerce sites evaluating automated customer returns, categorizing ML papers, and producing sentiment ratings for text.

The guide fine-tunes Qwen3.5 4B using datasets hosted on Hugging Face. Together AI says it samples 38,340 questions across six data sources to keep costs low, estimating about $17.0 in training costs, with larger datasets being more expensive and time-consuming. Setup involves cloning the tev1 GitHub repository, installing dependencies with uv sync --locked, and adding a TOGETHER_API_KEY to a .env file.

Training is launched through Together AI's fine-tuning service using a Python script, and Together AI says the job takes roughly 25 minutes. The resulting model is deployed to a dedicated endpoint via a CLI command using 1x_nvidia_h100_80gb_sxm hardware. Querying uses JSON input containing state, question, and labeled options; Together AI shows a charge-dispute example returning label "A" and key "duplicate_charge". The post notes that when calling the API directly or using Chat Playground, users should explicitly set temperature=0, max_tokens=8, and chat_template_kwargs={"enable_thinking": false}, because the public endpoint does not inject those defaults. Endpoints can be stopped with a CLI command and restarted later, or users can instead use together/Tev1-4B-experimental on the serverless platform.