Verified, then logged.
Sun 22 Nov 2026 · 00:46 KST / 15:46 UTC한국어 | ENSearch /

Breaking02 / IT

IBM's 20B slide model Palette arrives with no stated licence

IBM researchers report second place on the three-domain PresentBench mean, but their blog post does not say where the weights are distributed or what commercial terms apply.

A desk monitor in a dim meeting room displays a grid of wordless presentation slide thumbnails, one with overlapping shapes spilling past its edge. AI-generated image

IBM Research introduced Palette, a 20B slide-building model, on 3 Nov 2026 without stating a licence. Its researchers said in a Hugging Face blog post that the model is meant to run on users' own infrastructure.

Specification. The researchers said a single LoRA adapter, a lightweight fine-tuning method, serves two roles. The designer writes a JSON brief and the coder writes pptxgenjs code. Running the coder in parallel made the coding stage about 6–7 times faster.

Training data. The researchers said they used about 1,000 synthetic decks that follow the IBM Design Language. A large open teacher model, a bigger model whose output serves as training data, generated roughly 10,000 slides. Slides flagged for overlap or overflow were repaired and their reasoning traces rewritten.

Output first passes through a check-and-repair harness, a software layer that inspects model output, according to the blog post. An editor model fixes flagged slides, and a fix is kept only if it improves the slide. In the researchers' own measurement, the harness added an average of 3.5 points across 131 decks.

The researchers said Palette averaged 59.7 points across three domains of PresentBench, a benchmark for generated slide decks, placing second behind NotebookLM at 60.8. By domain, they put it first in Economics and Education and third in Talk.

The comparison figures come from the PresentBench paper, and a separate frontier model drafted the plans used in the evaluation, the researchers said. PresentBench does not measure adherence to a specific design language.

The blog post describes Palette as an open-weight model but does not give the location of the weights, commercial or redistribution terms, or minimum VRAM. The researchers said their next step is to use the harness's checks as reinforcement-learning rewards.