IBM Researchers Trained a Small Open AI Model to Build Slide Decks in Your Brand's Own Style

Key points

• Palette, a fine-tuned 20-billion-parameter open AI model trained on IBM's design language, ranked first on two of three domains in the PresentBench presentation benchmark.
• The system splits deck-making into two roles: a "designer" that plans the content and a "coder" that draws each slide, cutting the coding stage to roughly one-sixth of the time a single-pass approach takes.
• Training used roughly 1,000 synthetic slide decks (about 10,000 slides total), each checked and repaired by the same automated quality tool used at runtime.
• Palette runs on a company's own infrastructure, so confidential content never leaves internal systems.
• A plain-English directive system lets users steer individual slides, for example typing [[make this a timeline]] to switch a bullet list into a visual timeline.

Cloud AI services produce competent slide decks. Ask them to follow your company's exact brand guidelines while keeping sensitive content off external servers, and both requirements start to conflict with how those services are built. IBM researchers set out to close that gap with Palette, a presentation-building system published this week and first described by Hugging Face. It uses a fine-tuned open-weight model, one companies can download and run on their own hardware, trained specifically on the IBM Design Language.

Why is making a slide deck hard for a small AI model?

Small models struggle because slides require invisible arithmetic. The model doesn't draw anything directly. It writes code that places every word and shape at precise coordinates, calculating in its head whether a 62-character title will fit at a given font size without bumping into the chart below. Get that wrong and titles vanish into the background or text spills off the edge.

Palette tackles this by splitting the job the way a human designer actually works. A "designer" role reads your content and writes a structured plan: which slides get a chart, which get a table, and exactly what each one says. A "coder" role then takes one slide at a time from that plan and writes the code to draw it. Because each slide's instructions are fixed before any drawing starts, all slides can be coded in parallel, making that stage roughly six to seven times faster than doing them one by one.

Both roles run inside a single LoRA adapter (short for Low-Rank Adaptation, a technique that teaches a model a new skill without rewriting everything it already knows).

How does the quality-checking work?

Even a well-trained model occasionally produces a broken slide: two boxes stacked on top of each other, or a title the same colour as the background. Every finished deck passes through a checker before the user sees it. Simple rules catch known errors first. Then each slide is rendered and measured by a geometry checker that flags anything overlapping or running out of bounds. A separate editor model rewrites only the broken slides, and a fix is kept only if the checker confirms the problem is gone.

That same checker shaped the training data. The researchers generated around 1,000 synthetic decks, rendered every one, repaired broken slides, and turned each mistake into a better training example by recording what went wrong and why the fix worked.

When we covered IBM's CUGA agent framework on 3 November 2026, the finding was similar: a well-engineered wrapper around a modest model can outperform a much larger one on a bounded task. Palette makes the same argument for presentations, and it's a persuasive one. The benchmark lead is real, but it covers presentation generation specifically. Nobody should read it as a general challenge to frontier systems.

What does this mean for people who make presentations at work?

For most workers, nothing changes today. Palette is a research system, not a product on sale. It points toward a future where a company's IT team runs a private presentation tool that already knows the approved fonts, colours, and legal constraints about what content can leave the building. A branded draft in seconds, with nothing uploaded to an outside service.

Users who do get access can steer individual slides in plain English. Wrapping a short instruction in double brackets, such as [[lead with the ask]] or [[this is for the board]], tells the model to adjust that slide's layout or tone without touching the rest of the deck. The model follows directives because it was trained on them across twelve categories, each phrased in several ways, so it learns intent rather than a fixed phrase.

Common questions

Does Palette replace tools like PowerPoint or Google Slides?

No. Palette generates the code that fills a presentation file. You still open and edit the result in a normal slide tool; Palette just builds the first draft automatically.

Can any company use this right now?

Not as a finished product. The researchers published their method and benchmark results, but Palette is a research prototype. Companies wanting something similar would need to apply the same fine-tuning approach to their own design language and infrastructure.

What is PresentBench?

PresentBench is a benchmark, a standardised test designed to score AI models on how well they build slide presentations. Palette ranked first on two of its three measurement categories.