Together: Tev1 4B Experimental

Tev1 4B Experimental is Together AI’s classification model for questions with a fixed answer set. The manufacturer fine-tuned it from Qwen3.5 4B and reports 4.7 billion parameters. It uses non-thinking text generation to select an option: the documented input is a state, one question and labeled choices, and the requested output is exactly one letter. The application maps that letter to the option’s semantic key. This separates its decision workflow from a model that directly returns calibrated probabilities or an expected rubric score.

State, question and 2–24 options

The producer’s interface represents the decision as a JSON object with state, question and options. Each option contains a letter label from A through X, a semantic key and a description. Supply between two and 24 options and make each description explain when that answer applies. A letter is the compact output label; the semantic key is the meaning the surrounding software uses. Keep the mapping associated with the exact option list used for that decision.

For a support message, the state should contain the message and relevant context, while the question defines which intent or destination to choose. An illustrative list can distinguish a reported integration failure, an access issue and a product-information request, with a none option when the listed intents do not fit. Define overlapping cases rather than expecting the model to infer the organization’s routing policy from abbreviated labels.

Non-thinking generation and the requested answer shape

Together’s published example supplies a system instruction that treats content inside state as data, selects one listed option and requests only its letter with no explanation. Its recommended settings are temperature 0, an eight-token generation allowance and thinking disabled. The maker explicitly says these settings are not injected automatically by the public endpoint; the example helper supplies them, while a direct request must specify them. Use these settings to request the intended short, non-thinking classification response.

The eight-token allowance is an example output setting for this short-answer workflow, not a universal maximum output specification. The intended useful output remains the selected letter. If a helper exposes a label-and-key object afterward, distinguish that interpreted object from the model’s raw generated content. A category selection does not itself update a ticket, approve an operation or execute the next action.

Classification, rules and categorical ratings

The maker describes intent detection, policy application, workflow routing, document categories, sentiment and passage-based yes/no comprehension as supported task patterns. A boolean task can supply two options whose keys mean true and false; the application interprets the selected option accordingly. A sentiment task can supply described levels and return the key for one selected category. That categorical choice is not a probability-weighted average across levels, and a returned letter is not a numerical confidence value.

For a rule-based decision, include the relevant facts and the rule in the task rather than asking the model to invent the policy. For example, a document-routing brief can describe the permitted folders and the material that belongs in each. A research-taxonomy task can provide a paper’s title and abstract together with topic descriptions. The result should be interpreted against the supplied alternatives, with the original evidence retained for an ambiguous or disputed choice.

Training scope and edition identity

The model page records 37,840 training examples covering entailment, passage yes/no questions, banking intents, news categories, sentiment levels, programmatic policies, routing and research-paper taxonomy. The companion guide documents normalizing these tasks into a shared decision format and fine-tuning a classification model. This explains the profile’s focus on a bounded answer set; it does not establish that every new policy, domain or taxonomy will be handled correctly without task-specific assessment.

Tev1 4B Experimental is the named Together model, not the general Qwen3.5 4B base profile and not a checkpoint from another decision-model maker. Retain the experimental edition’s own output preparation and input contract. A common decision use case or a familiar base family does not justify borrowing another model’s image support, probability calibration, tool calls or question-batching limits.

Context and reviewing the result

Together’s model specification reports a 32.8K-token context window and text input with text output. The context has to accommodate the instructions, state, question and options used in the classification request. Prepare only the evidence the question needs, while retaining the definitions that distinguish the alternatives. Keep the classification prompt compact by removing irrelevant context while retaining the option definitions.

The producer identifies this as an experimental release intended for assessment on real decision workloads. Check that the raw response is one of the supplied labels before mapping it to a semantic key, and compare classifications with representative labeled material from the intended task. A valid letter establishes a well-formed selection, not proof that its interpretation is correct. Keep the input, options, returned letter and mapped key together so the final artifact can support routing, categorization or policy review without losing how the decision was defined.