PrismaCoder: Open 9B and 35B Models Close In on Proprietary Ones in Visual Code Generation

Open models for multimodal coding now come with three artifacts at once: the PrismaForge data pipeline, the PrismaCoder-2M dataset, and two models trained on it, PrismaCoder-9B and PrismaCoder-35B. The authors call PrismaCoder-2M the most diverse and largest open dataset for multimodal code. By their assessment, the models lead open solutions by a wide margin and compete with the leading proprietary systems. Code and data are published.

Why "code that looks right" is so hard to verify

Multimodal code connects the symbolic world of programs with the visual world they produce when run. Models are increasingly asked to generate charts, web pages, animations, and semi-structured documents. Closed systems can already build interactive artifacts from a text description or a single screenshot. But their data and training are not available, so the open community is stuck on the quantity and quality of training examples and on how to obtain them.

Quality is harder than quantity here. A good sample must be a correct program, be cleanly written, and, above all, look right once rendered. Existing signals cover these requirements only separately. Execution usually yields a single bit: pass or fail. Code similarity to a reference says little about the final image, since two completely different programs can draw the same thing, while nearly identical ones can produce different results. Task-specific verifiers do not transfer to other tasks.

A rubric as both filter and reward

PrismaForge is built as a hierarchical pipeline centered on a synthesized rubric. The rubric breaks the three quality factors down into verifiable criteria. The idea is that a general "good or bad" verdict becomes a set of concrete questions about a specific sample.

The same rubric plays two roles. First, it selects data for supervised training, forming the PrismaCoder-2M corpus. Then the same rubric acts as the reward in the reinforcement learning stage on visual coding tasks. This link removes a common problem: when the data filter and the RL signal measure quality by different standards, the model ends up optimizing for the wrong thing.

Another of the authors' results concerns transferability. Their analysis shows that the rubric-based pipeline improves multimodal coding across different base architectures. That suggests the effect comes from the approach to data and rewards itself, not from the quirks of a single model.

What exactly has been released

• PrismaForge: a hierarchical pipeline for data synthesis and filtering based on a decomposed rubric.
• PrismaCoder-2M: an open dataset for multimodal code that, according to the authors, is the most diverse and the largest to date.
• PrismaCoder-9B and PrismaCoder-35B: two models trained on this data, with weights released.
• Code and data are available on the project page.
• Task scope: charts, web pages, animations, and semi-structured documents.

What this means for developers and teams

The main value of the release is that the whole chain is open. Previously you could download a model but could not tell what it was trained on or how. Now the data, the filtering method, and the rubric are all available, and the rubric can be adapted to your own tasks. For teams building tools that generate dashboards, landing pages, or reports, this makes it possible to fine-tune a model on their own data with the same evaluation mechanism.

The model sizes also matter in practice. The 9B can realistically run on a single consumer or workstation GPU with quantization, which suits local scenarios and work with sensitive code. The 35B is aimed at more serious infrastructure, but it stays in a class that small teams can handle, not just labs with huge clusters.

Rubrics as a general-purpose signal can also be applied beyond visual code. Any task where the result must be judged along several independent axes, from interfaces to document generation, gets a template: break quality down into verifiable criteria, filter data with them, and use the same criteria as the reward. Claims of parity with proprietary systems should be checked on your own tasks, especially on complex interactive pages. Still, the open community now has a reproducible starting point.