🔍 Read the full analysis: 5 Ways Palette Teaches AI To Create On-Brand Decks on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Hugging Face describes Palette, a 20-billion-parameter open-weight model fine-tuned to generate slide decks in IBM Design Language. The company reports that it ranked first in two of three PresentBench domains and that parallel slide coding was about six to seven times faster than sequential coding; the supplied account gives no scores or independent comparison.
Hugging Face has described Palette, a 20-billion-parameter open-weight model fine-tuned to create slide decks in IBM Design Language, as detailed in the original analysis. The company says Palette ranked first in two of three PresentBench domains it measured, but the supplied account does not give scores, identify the domains or name the competitors, leaving the result difficult to assess independently.
Palette uses one LoRA adapter for two roles in deck creation: a designer and a coder. The designer turns a request into a structured JSON brief, setting out the deck’s content, palette, typography, slide layouts and design rationale. The coder then uses each slide’s specification to generate JavaScript for the pptxgenjs presentation library.
Hugging Face says separating those tasks lets slide coding run in parallel, making that stage about six to seven times faster than running it sequentially. That is a project-reported speed comparison; the supplied material does not describe the test setup or report an independent measurement. The company says coder examples were about five times longer than designer examples, so it oversampled designer examples fivefold during training.
Because the team says no public corpus contains the plans, design decisions and code behind finished decks, it synthesized training data with a larger open model. Hugging Face reports creating about 1,000 decks, covering roughly 10,000 slides, with varied genres, lengths, layouts and prompts. Examples also included counterfactual explanations of why a design choice was preferred over an alternative.
The team says it rendered generated decks and checked for overlaps and text overflow, repaired broken slides, and added revised examples to training. It also reports a content-faithfulness problem: when a specification lacked details, Palette could invent material. According to the account, reference-deck examples improved the teacher model’s output more reliably than adding further prompt rules.
A More Structured Route to Branded Slides
Presentation generation requires a model to do more than summarize information. It must decide what belongs on each slide and translate that plan into readable layouts, typography and shapes. Palette’s two-role workflow addresses those linked tasks by establishing a deck-wide plan before producing slide code, rather than asking one generation step to handle both editorial choices and geometry.
The approach may matter to organizations that want presentations to follow an existing design system. Hugging Face presents Palette as adaptable to a user’s design language, while the supplied project account focuses on IBM Design Language. A locally run open-weight model could also give organizations more control over where prompts and documents are processed. Those are potential advantages, not demonstrated deployment outcomes: the account reports no customer deployments and does not establish how Palette handles sensitive information in practice.
The reported benchmark ranking offers a limited signal that a smaller model can perform well on some measured deck-building tasks. Without domain names, scores, evaluation conditions or comparisons with other systems, it does not establish that Palette is broadly better than frontier models or ready for demanding business use.
Products Worth Considering
Jackery HomePower 3600 Plus Portable Power Station, 3584Wh, 3600W
Bose SoundLink Plus Portable Bluetooth Speaker, Wireless Outdoor Speaker, Up to 20 Hours Battery Life, Waterproof and Dustproof, Black
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Waterproof, Dustproof, USB‑C, Up to 12 hr Battery, Black
As an affiliate, we earn on qualifying purchases.
How Palette Separates Planning and Code
Hugging Face frames the project around two difficulties with general-purpose models used for presentations: their slides may not follow an organization’s visual identity, and sending confidential material to an external service may be unsuitable. The project account describes on-premises or otherwise locally controlled operation as a motivation, but does not provide deployment tests or privacy guarantees.
Its design separates editorial judgment from slide geometry. The JSON brief makes the planned content and visual decisions explicit before code generation. Hugging Face says this structure can be edited by a person or another agent. The team also used LoRA fine-tuning, which the account describes as a less expensive way to adapt the base model to the deck-building format and task.
Rendered-slide checks address visible failures such as overlapping elements and text that extends beyond its intended area. They do not, on the information supplied, establish that the system verifies the accuracy of claims, data or citations in a deck. The account’s admission that incomplete specifications could lead to invented content underscores the difference between fixing layout and establishing factual reliability.
“The designer handles the editorial judgment, and the coder handles the geometry.”
— Hugging Face, describing Palette’s workflow
Benchmark and Release Details Missing
The supplied account does not include PresentBench scores, domain names, baselines, competing systems or evaluation conditions. Readers therefore cannot quantify the reported ranking or determine how the evaluation compares with other presentation-generation tests. The six-to-seven-times speed claim is also not accompanied by detailed test conditions.
The account does not identify Palette’s 20B base model, provide a release date, or say whether model weights, code or evaluation materials are publicly available. It also does not provide independent testing, results for design systems beyond IBM’s, or performance data for complex data-heavy decks and strict factual tasks. The company’s description of local deployment does not specify security controls or operational requirements, and the frequency of unsupported content generation remains unreported.
Independent Tests Would Clarify Readiness
The supplied material does not announce a release date or say when model weights, code or fuller evaluation results will be available. Those details would allow developers and organizations to reproduce the workflow and test the claims against their own deck requirements.
Useful follow-up evidence would include scored comparisons across named PresentBench domains, tests against other models under the same conditions, and results on different design systems and longer presentations. Evaluation should also report how often generated slides contain layout faults, how reliably the checking process catches them, and how frequently the model adds unsupported facts. Until such information is available, Palette’s benchmark ranking and speed improvement remain claims reported by the project team.
Key Questions
What is Palette?
Palette is a 20-billion-parameter open-weight model that Hugging Face says it fine-tuned to generate slide decks in IBM Design Language.
How does its deck-generation workflow work?
A designer role creates a structured brief for the deck, including its content and visual system. A coder role then turns slide specifications into pptxgenjs JavaScript for the presentation.
What did Hugging Face report on PresentBench?
The company says Palette ranked first in two of three measured domains. The supplied account does not name the domains or provide scores, competitors or evaluation conditions.
Can Palette be used with confidential material?
Hugging Face presents locally controlled deployment as a motivation for the project, but the supplied material does not report deployment results or specify security safeguards. It does not establish how well Palette protects sensitive information in practice.
Are Palette’s model weights publicly available?
The supplied account does not say whether the weights or code have been released, or provide a release date. Their availability and licensing remain unclear.
Primary source: Hugging Face · via ThorstenMeyerAI.com
