• Bottom line: Wait. Palette's 59.7 PresentBench mean trails NotebookLM's 60.8 by 1.1 points, and the post links no weights, code or paper.
• A single LoRA adapter plays both designer (JSON brief) and coder (pptxgenjs per slide); parallel coder calls make the coding stage about 6-7x faster, per IBM.
• A check-and-repair harness keeps a rewrite only if a geometry checker confirms improvement; IBM reports an average gain of 3.5 points across 131 decks.
• Training data is about 1,000 teacher-generated decks (roughly 10,000 slides); invented facts were the hardest failure, and the post reports no faithfulness metric.
• PresentBench does not measure adherence to a specific design language, so the headline claim, building decks in your design language, is not directly benchmarked.

Bottom line: Wait. IBM Research's Palette, described in a Hugging Face post, generates slide decks in a chosen design language with a 20B open-weight model that an organization can host itself. On PresentBench it ranks second on the three-domain mean, 59.7 against NotebookLM's 60.8. The result is promising, but the post links no weights, code or paper, so nobody outside IBM can reproduce it yet.

How Palette builds a deck

Palette splits deck building into two roles served by one LoRA adapter, according to IBM Research. The designer reads a plan and writes a structured JSON brief covering palette, type system and per-slide layout. The coder then writes pptxgenjs JavaScript for each slide. Because the brief fixes every slide's context, coder calls run in parallel, which IBM says makes the coding stage about 6-7 times faster.

The training data is synthetic and small

The training set is about 1,000 decks, roughly 10,000 slides, generated by an unnamed large open teacher model because no public corpus existed. Designs follow the IBM Design Language, and each example includes counterfactual reasoning explaining a rejected alternative. IBM says reference decks, not prompt rules, curbed invented facts, because the teacher copied examples more than it followed instructions.

The repair harness is the most reusable idea

Palette renders each slide and measures overlap, clipping and overflow with a geometry checker. Flagged slides go to a separate editor model, and the rewrite is kept only if the checker confirms improvement. IBM reports this adds 3.5 points on average across 131 PresentBench decks. The keep-only-if-better rule makes automated repair safe by default and transfers to other generation tasks.

What the benchmark does and does not show

Palette placed first in Economics and Education and third in Talk, where two closed systems scored higher, per the post. PresentBench scores fundamentals, visual design, completeness, correctness and fidelity. The authors concede it does not measure adherence to a specific design language, so the central claim is untested. Per-domain scores are not given, and plan drafting relied on a frontier model.

Open gaps before anyone depends on it

Three gaps stand out. Faithfulness was the hardest failure, yet the post reports no measured rate of invented facts after the fix. Figure-heavy decks remain a known weakness, with figure extraction only planned. And the post links no weights, code or paper, so the results are a vendor-side report rather than something a reader can rerun and argue with.

The bottom line: Wait

Bottom line: Wait. The deciding number is 59.7 versus 60.8, a near-tie with a closed product, reported by the builder, without released artifacts. Revisit when weights and a reproducible harness appear, and when someone measures design-language adherence and fact fidelity directly. Teams with confidential decks and a house style should watch it, since the self-hosting rationale is sound.