A small decision head is as accurate as a large one
On the same base model, a 0.5M parameter head is as accurate as a 27M parameter head at half the inference time.
Researchers did a controlled study of typed decision models. These models read text and answer structured questions, such as yes or no checks, with a small part called a head. The team trains five heads with one protocol: the same base model, data, recipe and seed. On Qwen3-0.6B, each head gets 81.3 to 82.0 percent accuracy in domain.
What they found
The complexity of the head does not change accuracy. One prompt for all questions of a text gives the same accuracy as one prompt for each question. It gives 1.6 times faster inference and 2.1 to 2.5 times faster training on one H100.
The numbers
The size of the base model is more important. Qwen3-4B adds 4 points in domain and 11 points on held out sources, compared to Qwen3-0.6B. The three versions of one model give the same result.
What is not known
The authors use English public data and one seed. TypeSafe announced Jev in September 2026. Its structure is not published. The authors did not test it. They tell teams to make the base model better and keep the head small.
Insight: The base model is the real edge
TypeSafe started this type of model with Jev, which has an unpublished structure. Open copies came after it, such as Sarvam Prism and Microsoft Decision-1. Microsoft says that Decision-1 is 2.8 times as fast as Jev. This study seems to show that a head is easy to copy. The value is in the base model and the data. Judge vendors on those two.
Sources
Posted
