Palette lets a 20B open-weight model build decks in your design language, with a 59.7 three-domain mean 1.1 points behind NotebookLM

A self-hostable 20B open-weight model can now produce decks in a specified design language instead of inheriting a frontier model's default style. Frontier models produce good decks, but ten of them start to look alike in the same house style that prompting does not pull back to your template, while small open models that can actually be self-hosted previously produced broken decks with titles vanishing into the background, text running off the slide and shapes stacked on each other. On PresentBench, Palette ranks first on Economics and Education, third on Talk, and second on the three-domain mean at 59.7, 1.1 points behind NotebookLM (60.8) and ahead of the rest of the field.

The approach splits one generation into two kinds of decisions, editorial judgment and geometric placement, handled by a designer and a coder. A single pass has to decide what each of 10 to 20 slides says and how to show it while also placing elements at exact coordinates and font sizes, which small models struggle to do at once. The designer writes a structured JSON brief with the palette, the type system and, for each slide, its layout, content and a short design rationale; the coder turns one slide's spec into pptxgenjs JavaScript. Because each slide's context is fixed before any code is written, coder calls run in parallel, making the coding stage about 6 to 7 times faster.

Training data is synthesized by a large open teacher model and carries counterfactual reasoning about each choice. No public corpus holds plans, design decisions and the slide code that produced them, only finished decks. Recipes set each deck's genre, length, layouts and user directives, yielding around 1000 decks and roughly 10,000 slides; every example carries reasoning that weighs an alternative and says why it was rejected, and broken slides are repaired by the same checker used at inference with their reasoning rewritten.

At inference every deck passes through a check-and-repair harness, and a rewrite is kept only when the checker confirms the slide improved. Even a well-trained model sometimes produces a broken slide, such as a title in the same color as the background or two boxes stacked on top of each other. Simple rules fix known mistakes, then each slide is rendered and measured, a geometry checker flags overlapping, clipped or overflowing text, and a separate editor model rewrites that slide's code; across 131 PresentBench decks the harness adds 3.5 points on average.

Evaluation runs on PresentBench, which scores decks on fundamentals, visual design, completeness, correctness and fidelity and supplies a source document rather than a plan, so in these runs a separate frontier model first drafted each plan from its document and Palette built the deck from that plan; the report covers the Economics, Education and Talk domains, with baseline numbers taken from the PresentBench paper. The authors say they are not aware of a benchmark that measures whether a deck follows a specific design language.

The authors say the hardest problem, faithfulness, could not be stopped by a prompt rule when slides carried facts absent from the input, and was fixed by reference decks instead; on the narrative, speech-driven Talk domain two closed systems score higher and Palette places third. The authors also want to close the gap on figure-heavy decks and are exploring re-skinning output from one or two existing .pptx files without retraining.