IBM Palette

IBM Palette is a research system published by IBM Research on 2026-11-03. Its bet: a small, self-hostable open-weight model can build slide decks in a company's own design language (here, the IBM Design Language) and compete with frontier tools.

Palette fine-tunes a 20B open-weight base (the repo refers to gpt-oss-20b-palette-lora) with a single LoRA adapter that serves two roles:

• Designer – reads the plan and writes a structured JSON brief: palette, type system, and per-slide layout, exact content and design rationale;
• Coder – takes one slide spec and writes the pptxgenjs JavaScript that draws it.

A check-and-repair harness renders every slide, flags overlap/clipping/overflow, sends flagged slides to an editor model, and keeps a rewrite only if the checker confirms it improved. Output is a downloadable .pptx plus PDF preview.

What problem it solves

Enterprise decks suffer from two issues: frontier models produce a recognizable "AI house style" that ignores brand guidelines, and confidential content shouldn't leave the company. Palette uses an owned, self-hostable open-weight model to address both.

Key techniques

• Role split: the designer fixes the story and layout first, then the coder draws each slide; since each slide's context is fixed before coding, coder calls run in parallel, making the coding stage about 6–7× faster.
• Synthetic data: a large open teacher generated ~1,000 decks / 10,000 slides, each choice paired with counterfactual reasoning (e.g., why a chart instead of a table).
• Faithfulness first: IBM says the hardest problem was invented facts; changing the teacher's reference decks helped more than prompt rules.
• Check-and-repair: across 131 PresentBench decks the harness added 3.5 points on average.

Results and caveats

On PresentBench Palette ranks first on Economics and Education, third on Talk, and second on the three-domain mean (59.7 vs NotebookLM 60.8), ahead of Manus 1.6 and others. Caveats: results are self-reported; PresentBench measures general deck quality, not brand-language adherence; weight availability is not yet clearly stated.

Who it's for

• Enterprise design/marketing teams that need many decks in one brand system;
• Organizations that can't send content to external model servers;
• Engineers exploring the "small model + synthetic data + verifier harness" recipe.

Getting started

1. Read IBM Research's Hugging Face post Palette: teaching a small model to build decks in your design language for the designer/coder pipeline;
2. Watch IBM's project-palette repo: a chat-style harness where you describe a deck or upload PDF/DOCX/PPTX references → get an editable Markdown plan → the LoRA writes per-slide pptxgenjs → a Node renderer + LibreOffice produce .pptx and PDF;
3. Retry or refine individual slides, then download the .pptx.

Transferable lessons

Put the reasoning in the data, build the checker first, and keep a repair only when it measurably helps — applicable to any output you can check in code (charts, UI layouts). IBM next plans GRPO-style RL using harness signals as rewards.