IBM’s Palette shows a 20B open model can build on-brand slide decks nearly as well as NotebookLM

Slide decks sound trivial until a company wants every one on its own template, or the content is too sensitive for an outside model. Frontier systems drift back to their house style; small self-hosted models usually produce broken slides with vanishing titles, overflowing text and stacked shapes. Palette, from IBM Research, fine-tunes a 20B model with a LoRA adapter on synthetic decks following the IBM Design Language. On PresentBench it ranks first on Economics and Education, third on Talk and second overall, 59.7 against NotebookLM’s 60.8. If that holds, owning a deck model that speaks your brand becomes realistic for organizations that can’t send content to a frontier API, part of a wider run of capable small open models (apex-flash-1, Mellum2.1).

A fix that can’t make things worse can run on everything, which makes the harness the part worth copying for charts, diagrams and UI layouts. The team is candid about the hardest failure: slides containing facts never in the input. Prompt rules barely moved the teacher model; fixing the reference decks did, a sharp lesson about data over instructions. Hallucinated numbers in a board deck are real harm, and a geometry checker can’t catch them, so fact tracing, which the team names as a next signal, is where effort should go.

The training data is synthetic and generated by a large open teacher, so Palette’s taste is inherited. In benchmark runs, a separate frontier model drafted each plan from the source document before Palette built the deck. And PresentBench doesn’t measure adherence to a specific design language, the headline promise, as the authors acknowledge. Until that gap closes, “your design language” is a claim to test. The team plans to use checker signals as reinforcement learning rewards and is exploring a bring-your-own-template approach from existing .pptx files. More in our research coverage and tools coverage.