Together: Tev1 4B Experimental

Together: Tev1 4B Experimental is an experimental decision model from Together AI. It is a supervised fine-tune of Qwen3.5-4B, trained to choose one option from a structured state, a question, and a list of options. The model sits in the decisions category, and its name reflects its status: it is an experimental release rather than a general purpose chat model.

The model takes text as its only input and produces decisions as its only output. Its context window is 32768 tokens, which is enough to hold a structured state, a question, and a list of 2 to 24 candidate options in a single request. The maximum output is 8 tokens, which fits the shape of the task: the model returns a choice, not an essay. There is no free-form text generation here, and no image, audio, or other non-text input.

Because the output is a single decision drawn from a supplied list, the practical use cases are ones where a system needs to pick one option from a known set. Examples include routing a request to the correct handler, selecting a category or label from a fixed taxonomy, choosing the next action in a workflow, or resolving a branch in an agent loop. The structured state and question give the model the context it needs, and the option list bounds the answer space, so the result can be consumed directly by downstream code without parsing prose.

You call the model through api14's OpenAI-compatible API using the model id "togethercomputer/tev1-4b-experimental". Because the interface follows the OpenAI chat completions shape, you send your state, question, and options as text input in the usual request body and read the decision from the response. The model supports no additional request parameters beyond the standard ones, so there is no temperature, top_p, or similar sampling control to tune. With a maximum output of 8 tokens, responses are short by design.

This model fits decision tasks with a bounded set of options, not open-ended conversation, long-form writing, or tool-heavy reasoning. If your task needs generated prose, multi-step explanation, or inputs other than text, another model will fit better. The 32768 token context and the 2 to 24 option range also define the shape of problems it was trained for, so keep the state and question concise and the option list within that range.