togethercomputer/Tev1-4B-experimental

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026Architecture:Transformer0.0K Featherless Exclusive Cold

Tev1-4B-experimental is a 4.5 billion parameter experimental decision model developed by Together AI, based on a supervised fine-tune of Qwen3.5-4B. It is specifically designed to select one option from a structured state, question, and list of choices, retaining Qwen's standard next-token language-model head. This model excels at structured decision-making tasks, providing a single-letter output for automated classification and policy-transfer scenarios. Its 32768-token context length supports complex decision inputs.

Loading preview...

Overview

Tev1-4B-experimental is a 4.5 billion parameter experimental decision model from Together AI, fine-tuned from Qwen3.5-4B. It is designed for structured decision-making, where the model selects a single option from a given state, question, and a list of 2-24 choices. This model is an experiment inspired by the Jev concept, focusing on classification rather than non-autoregressive runtime.

Key Capabilities

  • Structured Decision-Making: Optimized to evaluate a decision task with a defined state, question, and options, returning a single letter corresponding to the chosen option.
  • High Accuracy on Specific Tasks: Achieved 88.0% on its main development decision set and 100% on a policy-transfer set during internal evaluation.
  • Controlled Output: Designed to return only a single letter, with no additional explanation, making it suitable for automated systems.
  • Qwen3.5-4B Base: Leverages the underlying architecture of Qwen3.5-4B, maintaining its standard next-token language-model head.

Good For

  • Automated Classification: Ideal for scenarios requiring a definitive choice from a predefined set of options based on provided context.
  • Policy-Transfer Applications: Demonstrated strong performance in synthetic policy-transfer tasks.
  • Integration with Application Logic: Its single-letter output is intended to be mapped back to semantic keys by application code, simplifying integration.

Limitations

  • Not intended for generic chat or prose generation.
  • Should not be used as the sole authority for high-impact decisions due to potential inaccuracies.
  • Aspects like prompt injection, multilingual behavior, and out-of-distribution robustness have not been comprehensively evaluated.

Finetunes
2 models