IBM Research said on November 3 that it fine-tuned a 20-billion-parameter model, Palette, to generate presentation slides in the IBM Design Language, and reported it ranks second overall on PresentBench.
The model is aimed at organizations that want decks in their own visual style without sending content to a frontier model's servers. PresentBench, however, scores general deck quality, not adherence to a design language. IBM says it is not aware of a benchmark that measures that, so the reported numbers do not test the model's central purpose. The results are IBM's own, in a Hugging Face blog post; baseline scores are taken from the PresentBench paper and have not been independently replicated.
Why IBM built it
According to IBM's blog post1, frontier models drift back to their own house style regardless of prompting, and some content cannot leave a company's systems. IBM's bet is that a small open-weight model trained on synthetic data and checked by a harness can compete with frontier systems while the user keeps ownership of the result.
IBM says small models are poor at slide generation for three reasons. They write code that places each element by coordinates and font size without seeing the result, so overlapping shapes and text running off the slide are common. A single pass must also handle editorial choices and precise placement across 10 to 20 slides. And IBM says no public corpus exists that shows the reasoning behind design choices or the code that produced them.
Designer and coder roles
IBM describes Palette as splitting the work the way it says people build decks. Two roles are served by a single LoRA adapter. The designer reads a plan and writes a structured JSON brief covering palette, type system, and each slide's layout, content and design rationale. The coder takes one slide's specification and writes JavaScript for the PptxGenJS2 library that draws it.
Because each slide's context is fixed before coding starts, coder calls run in parallel, which IBM says makes the coding stage about 6 to 7 times faster. Users can also steer individual choices with plain-English directives in double brackets, such as [[make this a timeline]]. IBM says the training data covers directives in 12 categories.
Training data and the repair harness
IBM generated about 1,000 decks covering roughly 10,000 slides using a large open teacher model. Each example carries a counterfactual reasoning trace that weighs an alternative and says why it was rejected. IBM said its largest problem was faithfulness, meaning slides that contain facts absent from the input. It said changing the reference decks shown to the teacher worked better than adding prompt rules.
Every deck also passes through a check-and-repair harness before it is returned. Simple rules fix known errors such as malformed color codes. A geometry checker then renders each slide and flags overlap, clipping or overflow. Flagged slides go to a separate editor model, and a rewrite is kept only if the checker confirms an improvement. IBM generated each PresentBench deck once and scored it with and without the harness; across 131 decks the harness added 3.5 points on average.
What the benchmark comparison shows
PresentBench scores decks on fundamentals, visual design, completeness, correctness and fidelity. It supplies a source document rather than a plan, so for these runs a separate frontier model, which IBM does not name, drafted each plan and Palette built the deck from it. The scores therefore reflect that whole pipeline, not Palette alone.
IBM compared Palette with NotebookLM, Manus 1.6 and five other systems. Palette ranks first on the Economics and Education domains and third on Talk, where decks are narrative and speech-driven and two closed systems score higher. Across the three domains it placed second, 1.1 points behind NotebookLM (59.7 versus 60.8). IBM describes Palette as a single 20B open-weight model.
The post does not identify the 20B base model or the teacher model. It also does not say whether or where the weights have been released, and a GitHub repository for the project could not be accessed.
IBM said its next steps are to use the harness's checks, such as whether a slide renders and whether anything overlaps, as rewards for reinforcement learning so the model makes fewer errors at generation time. It also plans to extract figures from source documents for figure-heavy decks and is exploring a way to apply a user's own templates from one or two existing .pptx files without retraining.