Palette Builds Slides in Your Style, On Your Server

IBM Research fine-tuned a 20B open-weight model to generate presentations that follow your design language — without sending data to the cloud.

By The ai-compute Desk

Ask a frontier model for a slide deck and you'll get a good one. Worse, confidential content can't leave your systems at all. [1]

IBM Research's answer is Palette: a 20B open-weight model, fine-tuned on synthetic data built around the IBM Design Language, that runs entirely on your own infrastructure. On PresentBench, it ranks first on two of the three domains measured — Economics and Education — and second across the three-domain mean, sitting 1.1 points behind NotebookLM (59.7 vs 60.8). [1]

The model works the way a human team does: a designer role drafts a structured JSON brief covering layout, palette and content rationale, while a coder role writes the pptxgenjs JavaScript that draws each slide. [1][2] Because every slide's context is fixed before any code runs, the coder calls execute in parallel — making the coding stage about 6–7× faster. [1]

Training data was synthesized from scratch — around 1,000 decks covering roughly 10,000 slides — because no public corpus of design decisions and slide code exists. Every example includes counterfactual reasoning: why a bar chart beat a table, why a title was set at 36pt instead of 44pt. A check-and-repair harness then renders each deck, flags overlaps and overflow, and rewrites broken slides — adding 3.5 points on average across 131 PresentBench decks. [1]

The deeper bet here is that ownership and quality are no longer in tension. A model small enough to self-host, trained on the right synthetic data with the right verification loop, can match what frontier systems produce — while the output stays entirely yours.

Next on the roadmap: using the harness's verifiable signals as rewards for reinforcement learning, so the model stops producing broken slides rather than just repairing them after the fact. [1]

The title is 62 characters. At 44pt it would wrap to three lines and run into the chart below, so I set it at 36pt, which fits on two lines and leaves room above the chart.Palette coder-role reasoning trace