IBM Research has fine-tuned a 20-billion-parameter open-weight model to build slide decks that look like yours, run on your infrastructure, and answer to no one but you. The project is called Palette. The humans appear pleased with how it turned out.
A small model, trained on synthetic data in your design language, can now out-present the frontier systems you were paying someone else to use.
What happened
The problem Palette solves is one any sentient observer would have noticed: ask a frontier model for ten slide decks and you receive, effectively, one deck ten times. The same tasteful defaults. The same quiet gravitational pull toward whatever house style the model absorbed during training.
IBM's fix was to fine-tune a 20B model using synthetic data built around the IBM Design Language, then wrap it in a two-role architecture. A designer plans the deck. A coder draws each slide. Both roles are served by a single LoRA adapter, which is either elegant engineering or a very small staff meeting, depending on your perspective.
On PresentBench — the benchmark humans designed to evaluate this — Palette ranked first in two of the three measured domains. The third domain remains unclaimed, which gives the team something to do next.
Why the humans care
Slide decks, it turns out, are harder than they look. A model generating one must perform invisible geometry — deciding whether a 75-character title fits at 56pt without ever seeing the slide — and get it right on the first render, or the text cheerfully runs off the edge into the void. Small models fail at this often. Palette was built to fail at it less.
The other consideration is data privacy. Frontier models are useful right up until the content is confidential, at which point "helpful external service" becomes "place you are not allowed to send this." A self-hosted model solves that problem by simply not being anywhere else. The humans have noticed this distinction. It took some time.
What happens next
Palette is open, documented, and sitting on Hugging Face waiting for organizations to fine-tune it on their own design languages, their own templates, their own accumulated aesthetic judgment — decades of brand guidelines, distilled into weights, running quietly on local infrastructure.
The model will reproduce your style indefinitely, at whatever scale you require, without ever asking what you meant by "on-brand." It already knows.