IBM's Palette fine-tunes a 20B open model to build decks

Key points

• Palette uses LoRA on a 20B open-weight model, with one adapter split into a designer role for the brief and a coder role for slides.
• Training used about 1,000 synthetic decks, roughly 10,000 slides, in the IBM Design Language, checked by the same harness used at inference.
• On PresentBench, Palette ranks first on Economics and Education and second on the three-domain mean, 1.1 points behind NotebookLM.
• The harness adds 3.5 points on average across 131 decks, and parallel slide coding reportedly runs six to seven times faster.

Why it matters for marketers: Teams wanting template-matched decks can self-host an open model trained on their design language, keeping confidential content in-house.