IBM Palette adapts a small model to a company’s presentation design language

What is changing?

IBM Research published Palette through Hugging Face on November 3. The system customizes a 20-billion-parameter open-weight model with LoRA so organizations can run it on their own infrastructure. The goal is to produce decks that follow company color, typography and layout rules without sending confidential content outside.

Palette gives the same model two roles. A designer prepares a structured brief covering content and visual decisions, while a coder draws each slide with pptxgenjs. Slides are generated in parallel, making the coding stage about 6–7 times faster. Training used roughly 1,000 synthetic decks covering around 10,000 slides.

Outputs are not returned directly. A harness measures overflow and overlap, and a separate editor keeps a repair only when the checker confirms an improvement. Across 131 decks, the harness added 3.5 points on average. Palette ranked first in PresentBench’s Economics and Education domains and second on the three-domain mean.