Physix FrontierPhysix Frontier · News Briefing Card (HuggingFace · Nov 4, 2026)
IBM fine-tunes 20B model on synthetic data for on-brand slides
KEY FACTS
- IBM Research released Palette, a 20B open-source model fine-tuned specifically to generate slide decks that follow corporate design languages.
- Palette splits the task into two roles, a designer and a coder, which share a single LoRA adapter.
- The training data was synthesized by large models: about 1,000 presentations and 10,000 slides, including design reasoning traces.
- On PresentBench, Palette ranked first in two of the three evaluated domains.
- During the coding stage, slides are processed in parallel, delivering roughly a 6x to 7x speedup.
KEY DATA
20BModel parameters
about 1,000Synthetic presentations
about 10,000Synthetic slides
6-7xCoding-stage speedup
PHYSIX OBSERVATION
A small model, powered by synthetic data and an editable intermediate representation, matches frontier models under specific design specifications, and its data never leaves the internal network. The signal is clear: the competitiveness of enterprise AI is shifting from model scale to private design languages and data closed loops, and self-hosted solutions are becoming practical.
Source: HuggingFace report