Shanghai AI Lab's Intern-Decision 4B edges Jev at 90.02%

Shanghai AI Lab's new multimodal decision models hit 90.02% unweighted average accuracy across seven evaluation suites, edging TypeSafe's Jev v1.13.0 at 88.74%. The paper on arXiv, posted October 15 by Tong Zhu, Zhe Xu and co-authors drawn from the lab plus Shanghai Jiao Tong University, Fudan University, and Nanyang Technological University, releases the family in three sizes: 0.8B, 2B and 4B, built on Qwen3.5 backbones.

The trick is architectural. Rather than generate structured output as free-form JSON, the model places a reserved placeholder token into an assistant JSON skeleton and reads candidate logits off the native language-model head in a single forward pass. In the authors' words: "Ground-truth symbols appear only in supervision, while all fields are evaluated in a single causal forward pass using the backbone's native language-model head."

On a single RTX 4090, a 289-token request with three decision fields runs 44.16 ms for the 4B and roughly 33-34 ms for the smaller models. Temperature scaling cuts expected multiclass Brier from 0.628 to 0.550 and expected calibration error from 0.213 to 0.089 on a 96-case known-distribution pilot.

A separate handoff study buries a sharper number than the headline. On 111 public Jevbench-Hard items, routing 42 low-confidence requests to a larger Atria-Dawn-Preview lifts accuracy from 73.87% to 87.39%. The release lands alongside AI Weekly's coverage of TypeSafe's Jev adoption: where Jev shipped as a typed-decision product behind an API, Intern-Decision ships as open weights.