Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo fine-tuned a 20B model on synthetic decks following IBM's visual system, then used a geometry checker to keep a repair only when the rendered slide improved.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [Hugging Face Newsroom](https://huggingface.co/blog/ibm-research/palette-deck-building)
Why it matters #
Palette treats enterprise design consistency as a training and verification problem. A self-hosted 20B model that plans, codes and repairs each slide could let companies keep confidential material inside their own systems while still getting decks that look like their own, provided the design-language claim holds beyond IBM's template.
Praveen Venkateswaran, Anupama Murthi and Hamid Adebayo of IBM Research published Palette on November 3rd, 2026, describing a 20B open-weight model fine-tuned on synthetic slide decks built around the IBM Design Language. The practical claim is narrow and testable: train a smaller model on one visual system, give it an intermediate plan to edit, and check each rendered slide before accepting a repair.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.