IBM Research's Palette: a 20B open model that builds slide decks

IBM Research published a post on Hugging Face on 3 November 2026 describing Palette, a 20-billion-parameter open-weight model fine-tuned to build slide decks in a fixed design language, here the IBM Design Language. IBM says it ranks first on two of the three domains of PresentBench, a public deck-generation benchmark, and second overall as a single self-hostable model.

Why IBM built it

The authors give two reasons a team might not want a frontier model for decks. Frontier models drift back to their own house style however you prompt them, and some content is confidential and cannot leave your systems. IBM's bet is that a small open model, trained on synthetic data in your design language and checked by a harness, can compete while you keep ownership of the result.

A designer and a coder

Palette splits the work into two roles served by one LoRA adapter (a small set of added weights, rather than a full retrain). The designer reads the plan and writes a structured JSON brief: palette, type system, and for each slide its layout, exact content and a short rationale. The coder takes one slide's spec and writes the pptxgenjs JavaScript that draws it.

IBM says the brief is readable and editable, so a person or an agent can change one slide without regenerating the deck. Because each slide's context is fixed before any code is written, the coder calls run in parallel, which the authors say makes that stage about 6 to 7 times faster.

Training data and what went wrong

IBM generated roughly 1,000 decks covering about 10,000 slides with a large open teacher model, because no public corpus of deck plans, design decisions and slide code exists. Every synthesized deck was rendered and checked for overlap and overflow, and broken slides were repaired and their reasoning rewritten.

• Faithfulness was the hardest problem. The authors say a prompt rule could not stop the model inventing content when a brief asked for, say, a six-by-five table but carried three bullets.
• What worked was fixing the reference decks in the teacher's prompt, which it copied far more than it followed instructions.
• Coder examples were about 5 times longer than designer examples, so the designer was undertrained until IBM oversampled its examples 5 times.

The harness

Output is not returned directly. Simple rules fix known mistakes, such as a malformed colour code that turns every slide black. Each slide is then rendered and measured by a geometry checker for overlapping, clipped or off-slide text. Flagged slides go to a separate editor model that rewrites the slide's code, and the rewrite is kept only if the checker confirms an improvement. IBM says this means the harness can fix a slide but cannot make one worse.

What the benchmark shows, and what it leaves out

PresentBench scores decks on fundamentals, visual design, completeness, correctness and fidelity. IBM reports Palette first on Economics and Education, third on Talk (two closed systems score higher), and second on the three-domain mean at 59.7, against 60.8 for NotebookLM. Baseline numbers come from the PresentBench paper. A separate frontier model drafted each plan from the source document before Palette built the deck, so the result depends on that planning step.

IBM notes that PresentBench does not measure whether a deck follows a specific design language, and says it knows of no benchmark that does. The post does not report latency, the hardware used, the cost per deck, the weights' licence, or how Palette copes with a template it was not trained on. IBM lists template re-skinning from existing .pptx files, and reinforcement learning from the harness's checks, as future work.

What a team can take from it

The reusable idea is the harness, not the model. If an AI job produces output you can check in code, such as a document that must render, a form that must validate or a layout that must not overflow, put a checker after the model and accept a rewrite only when the check improves. That makes the repair step safe to run on every output.