IBM Research's Palette builds slide decks with a 20B open-weight model and one LoRA adapter, scoring 59.7 on PresentBench versus 60.8 for NotebookLM
A small model plus a render-and-check harness gets close to a large product on deck generation, in a company's own design language.
// Key points
- One LoRA adapter plays two roles: a designer that writes a JSON brief per slide, and a coder that writes pptxgenjs per slide in parallel, about 6 to 7 times faster
- A harness renders each slide and checks overlap and overflow; a rewrite is kept only if the checker confirms it improved, adding 3.5 points on average across 131 decks
- Trained on about 1,000 synthetic decks (roughly 10,000 slides) from a large open teacher model, aligned to the IBM Design Language
- PresentBench: first on Economics and Education, third on Talk, second on the three-domain mean
Builder's takeFor PandaClaws' generated assets this is the pattern I'd copy: split plan from rendering, check the rendered output mechanically, and keep a fix only when the checker says it's better. The 3.5-point gain from the harness alone is a cheaper win than a bigger model.
// Background · from #Open models
Full timeline →- Nov 5 Hugging Face proposes Model Integration Contracts for Transformers: tiny real-architecture fixtures on CPU that would have caught a UMT5 bug that sat in main for almost two months
- Nov 4 Amazon Bedrock adds in-country inference in India for Kimi K3, keeping processing within the Mumbai and Hyderabad Regions
- Nov 3 AWS distills a GLM-5.3 shopping agent into Qwen3.5-4B: held-out success rises from 0.297 to 0.541, level with the teacher