Key points
- IBM Research fine-tuned a 20B-parameter open-weight model to generate slide decks adhering to the IBM Design Language.
- The system splits deck creation into two roles: a designer that writes JSON briefs and a coder that generates rendering code.
- Palette ranks first on two of three domains in PresentBench, a benchmark for presentation quality, according to the authors.
according to the published report, a 20B-parameter model can produce slide decks consistent with corporate design languages, according to a blog post published by Hugging Face.reported The model, named Palette, addresses the tendency of frontier systems to revert to generic styles and the inability of small open models to handle complex layouts.IBM Research created a model that makes professional-looking slide decks while keeping your private data safe on your own servers.IBM Research fine-tuned a 20B-parameter model with a two-role LoRA adapter to achieve state-of-the-art results on PresentBench for design-consistent deck generation. It operates on self-hosted infrastructure, keeping confidential content within user systems, while maintaining ownership of the generated output.
Evidence: reported
Palette: teaching a small model to build decks in your design languageaifleet-huggingface
The figures
The underlying model has 20 billion parameters, according to the IBM Research post.reported Training used approximately 1,000 synthetic decks, covering roughly 10,000 slides, as stated in the documentation.reported The coding stage runs 6 to 7 times faster than sequential generation because slides are processed in parallel, per the authors. Coder training examples were five times longer than designer examples, requiring a five-fold oversampling of designer data to balance the mix, according to the report.reported
Evidence: reported
Palette: teaching a small model to build decks in your design languageaifleet-huggingface
Evidence: reported
Palette: teaching a small model to build decks in your design languageaifleet-huggingface
Behind the figures
The team measured performance using PresentBench, ranking first on two of the three domains tested, according to the blog post.reported Training data was synthesised by a large open teacher model, starting from recipes that defined genre, length and layouts. This method allowed control over data coverage and included plans ranging from tidy outlines to raw content dumps, ensuring the model learned to condense information, as described in the material.
What changes, what does not
Palette separates editorial judgment from geometric placement by using a single LoRA adapter for two roles, according to the authors.reported The designer role produces a structured JSON brief containing palette, type system and layout decisions, while the coder role writes code to place elements at specific coordinates. This separation allows a person or agent to edit individual slides without regenerating the entire deck, a capability not present in single-pass frontier models, per the post.
Evidence: reported
Palette: teaching a small model to build decks in your design languageaifleet-huggingface
What builders can do with it
Developers can build self-hosted presentation tools that enforce strict design language compliance, according to the IBM Research post.reported The documented interface allows integration with existing design systems by feeding the model synthetic data in a specific format. Builders can implement parallel slide generation for faster output and enable human-in-the-loop editing of the intermediate JSON brief, without relying on external APIs for confidential content, as the material describes.
Evidence: reported
Palette: teaching a small model to build decks in your design languageaifleet-huggingface
Why it matters
- Policy readers should note that self-hosted models allow organisations to keep confidential content within their own infrastructure, avoiding external data exposure.
- The separation of design and coding roles creates an editable intermediate format, enabling auditability of design decisions before final rendering.
- Performance on PresentBench suggests small models can compete with frontier systems on specific tasks when trained with targeted synthetic data.
- The use of LoRA adapters indicates that fine-tuning for specific formats is cost-effective and preserves base model capabilities.
Topics: models, IBM, Hugging Face, synthetic data, fine-tuning
Updated
Then / now / next
Small open models produced broken slides with overlapping shapes and missing titles, according to the IBM Research post.
Palette uses a two-role LoRA adapter to generate consistent decks on self-hosted infrastructure, per the authors.
Developers may test whether this approach generalises to other design languages beyond IBM's, as no such results are reported.
Story timeline
- Then Small open models produced broken slides with overlapping shapes and missing titles, according to the IBM Research post.
- 03 Nov 2026 Palette: teaching a small model to build decks in your design language
- Now Palette uses a two-role LoRA adapter to generate consistent decks on self-hosted infrastructure, per the authors.
- Next Developers may test whether this approach generalises to other design languages beyond IBM's, as no such results are reported.
The strongest doubt
Counterpoint from the sources
The model was trained only on the IBM Design Language, so its ability to replicate other corporate styles remains untested, according to the post.
For readers
Policy
- Self-hosting keeps confidential data within organisational boundaries, reducing compliance risks.
- Editable JSON briefs allow for audit trails of design decisions.
Developers
- Parallel slide generation speeds up the coding stage by 6-7x.
- LoRA adapters enable cost-effective fine-tuning for specific formats.
Business
- Consistent branding is maintained without relying on generic AI styles.
- Ownership of the output remains with the user, not the model provider.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Terms in this story
LoRA adapter
A low-rank adaptation technique for fine-tuning models without altering the base weights.
PresentBench
A benchmark used to measure the quality of generated presentation decks.
JSON brief
A structured document containing design decisions and layout instructions for slides.
Synthetic data
Artificially generated training data created by a teacher model to simulate real-world scenarios.
Questions readers ask
Can Palette generate decks for any design language?
The model was trained specifically on the IBM Design Language, according to the blog post, so results for other languages are not documented.
How does Palette handle layout errors?
It uses a two-step process where a designer plans the layout in JSON and a coder places elements, reducing errors like overlapping shapes, per the authors.
Is the model available for download?
The blog post describes the training and performance but does not state whether the weights are publicly available for download.
Connected stories
Reading receipt
- IBM Research fine-tuned a 20B-parameter open-weight model to generate slide decks adhering to the IBM Design Language.
- The system splits deck creation into two roles: a designer that writes JSON briefs and a coder that generates rendering code.
- Palette ranks first on two of three domains in PresentBench, a benchmark for presentation quality, according to the authors.