IBM Research has released Palette, a 20-billion-parameter open-weight super intelligence (SI) model designed to generate slide decks that adhere to specific design languages, addressing the tendency of frontier SI systems to default to a uniform house style.

What Happened

Palette is built on a small open-weight SI model fine-tuned to mimic the workflow of human deck creation, splitting the task into two roles: a 'designer' that plans the content and a 'coder' that writes the code to render each slide. The model was trained on synthetic data generated by a larger teacher model, following the IBM Design Language. To ensure reliability, the team implemented a check-and-repair harness that validates layout geometry and fixes overlaps or text overflow before returning the final deck. The training process involved generating approximately 1,000 decks covering roughly 10,000 slides, with each example including counterfactual reasoning to teach the model why certain design choices were made or rejected.

Why It Matters

For developers and enterprises seeking to self-host SI tools, Palette offers a solution to the 'black box' nature of frontier models, where outputs are often confidential or stylistically constrained. By using a LoRA adapter, the system maintains the base model's general capabilities while specializing in deck generation, allowing it to run on owned infrastructure. The approach demonstrates that smaller SI models can compete with larger closed systems in specific domains when paired with rigorous evaluation harnesses. On the PresentBench evaluation, the company reports that Palette ranked first on two of the three measured domains (Economics and Education) and second overall on the three-domain mean, placing it ahead of several other open and closed systems.

The Bottom Line

Palette demonstrates that specialized, open-weight SI models can produce high-quality, design-consistent outputs by combining synthetic training data with automated repair harnesses, offering a viable alternative to frontier systems for structured content generation.