IBM Research has fine-tuned a 20-billion-parameter open-weight model, under a project called Palette, to build slide decks in IBM's Design Language. The team argues that small, self-hosted models are poor at slide layout and often produce titles that vanish into the background, text that runs off the slide, and overlapping shapes. Frontier models, by contrast, produce polished decks but drift toward their own house style and cannot be kept on private infrastructure. Palette tests whether a small model trained on synthetic data in a company's design language, and checked by a rendering harness, can compete while the organisation keeps ownership of the output. On PresentBench, the team reports that its model ranks first on two of the three domains it measured. The system splits deck creation into two roles served by a single LoRA adapter. A designer pass plans the story and writes a structured JSON brief covering the palette, type system, and each slide's layout, content and design rationale. A coder pass then writes pptxgenjs JavaScript for each slide. Because every slide's context is fixed before coding starts, the coder calls run in parallel, which the team says makes that stage about six to seven times faster. The training data was also synthesised. Around 1,000 decks covering roughly 10,000 slides were generated with a large open teacher model, and each example includes reasoning that weighs and rejects alternative layouts. Every synthetic deck was rendered and checked for overlap and overflow, and broken slides were repaired into new training examples. The hardest problem, the team says, was faithfulness: when a brief asks for more content than the source supplies, the model tends to invent it. Prompt rules did not stop this. Fixing reference decks did. For Australian businesses and professionals, this changes little today. Palette is a research write-up, and the article gives no licence terms, pricing or general release date, so it is not a product to buy yet. It may matter for organisations that want on-brand decks and cannot send confidential material to overseas cloud services, since self-hosted open weights make that possible in principle. Self-hosting does not remove legal duties, however. Any deck containing personal information remains subject to the Privacy Act 1988 and the Australian Privacy Principles. Teams already running open models internally could test the approach. Everyone else can watch for a usable release.