PrismaCoder: open 9B and 35B models target closed visual-coding lead

A new arXiv paper releases two open-weight models, PrismaCoder-9B and PrismaCoder-35B, for the slice of coding where the output is visual: charts, webpages, animations and what the authors call semi-structured documents. The paper's claim is that the two models "lead the open state of the art by a wide margin and rival leading proprietary systems." The abstract does not name which proprietary systems, and reports no benchmark score.

The method is pitched as a fix for a judging problem. A good visual-coding sample has to be "correct as a program, well built, and above all visually right once rendered," and the authors argue no existing training signal covers all three: "Execution often returns just a pass or fail, code similarity does not map to the visual output, and task-specific verifiers do not generalize." Their answer, PrismaForge, is a synthesized rubric that decomposes those three factors into checkable criteria and plays two roles — it filters the supervised corpus, then becomes the reward model for reinforcement learning on visual-coding tasks.

The accompanying dataset, PrismaCoder-2M, is described as "the most diverse open dataset for multimodal code intelligence and the largest to date." The paper is marked work in progress; until independent reproductions land, the parity claim rests on the authors' own evaluation.