Together AI has launched together/Tev1-4B-experimental, a classification model built on Qwen3.5 4B and served on its platform, alongside a step-by-step guide to fine-tuning a comparable model.
The model takes a block of state (for example, a customer support message) plus a fixed set of multiple-choice options and returns a single label — a score, boolean, or letter choice — rather than free text. Together's example walks through building a support-intent classifier: given a message about a duplicate subscription charge, the model correctly returns the label for "duplicate_charge" from a four-option list.
The guide details the training pipeline. A fine-tuning dataset of 37,840 examples is sampled from six public sources — MultiNLI, BoolQ, Banking77, AG News, SST-5, plus internal programmatic-policy and routing datasets — covering intent classification, yes/no comprehension, sentiment, and rule-based decisions. Training this dataset against Qwen3.5 4B costs about $17 and takes roughly 25 minutes on Together's fine-tuning service, after which the model is deployed to a dedicated HTTP endpoint (the example uses a single H100 GPU).
Together specifies exact inference settings needed to reproduce deterministic output: temperature=0, max_tokens=8, and enable_thinking=false, plus a system prompt instructing the model to treat the state field as data rather than instructions and to return only the selected option's letter.
Companies that don't want to train their own version can query together/Tev1-4B-experimental directly on Together's serverless platform without running the fine-tuning steps themselves.