IBM's Palette: a 20B model that builds on-brand decks

What happened?

IBM Research introduced Palette, a model fine-tuned from a 20B open-weight base that builds slide decks in the company's own design language (Hugging Face). The goal is a model you can self-host that doesn't drift back to the recognizable house style of frontier models, and that keeps confidential content inside your own systems.

Small open models have been bad at this: titles vanishing into the background, text running off the slide, shapes stacked on top of each other. The model never sees the slide. It writes code that sets coordinates and font sizes, and only finds out whether a title fits after rendering.

How does it work?

The work is split into two roles served by a single LoRA adapter. The designer turns the plan into a structured JSON brief: palette, type system and, per slide, layout, exact content and a short rationale. The coder takes one slide spec and writes the pptxgenjs JavaScript that draws it. Coder calls run in parallel per slide, which makes the coding stage about 6 to 7 times faster.

The training data is synthetic: a large open teacher model generated about 1,000 decks with roughly 10,000 slides, each choice paired with counterfactual reasoning (for example, why a chart and not a table). The hardest problem was faithfulness, keeping invented facts off the slides. Prompt rules barely helped. Fixing the reference decks did, because the teacher copied those far more than it followed instructions.

Users steer the output with directives in [[ ]], such as [[make this a timeline]] or [[no pricing]], drawn from 12 categories in the training data. A check-and-repair harness renders each slide, flags overlap and overflow, and hands broken slides to a separate editor model. A rewrite is kept only if the checker confirms it improved. Across 131 decks, the harness added 3.5 points on average.

What do the benchmarks show?

On PresentBench, Palette ranks first on Economics and Education and third on Talk. On the three-domain mean it places second with 59.7, 1.1 points behind NotebookLM (60.8) and ahead of Manus 1.6 and five other systems. Plans were drafted from source documents by a separate frontier model, and Palette built the decks from them.

Source: IBM Research. Palette (blue) PresentBench scores by domain, next to NotebookLM, Manus 1.6 and five other systems.

What does it mean in practice?

The post doesn't say the weights or adapter are available to download, so for now this is a recipe rather than a ready tool. The recipe travels well, though: put the reasoning in the data, build the checker first, and keep a repair only when it measurably helps. That applies to any output you can check in code, from charts to UI layouts. Next, IBM wants to use the harness signals as rewards for GRPO-style reinforcement learning.