Is Palette state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
IBM Research's 20B slide-deck model, trained as a single LoRA adapter: a designer role writes a JSON design brief and a coder role writes PptxGenJS code for each slide in a company design language, with a check-and-repair harness. No checkpoint or API published yet.
The details so far
- Lab
- IBM
- Announced
- Nov 3, 2026
- API
- Not yet · None announced
- Open weights
- Unknown · Described as open-weight, but no checkpoint, Space or code repo was published as of Nov 3, 2026
- Parameters
- 20B (per IBM Research; base model not named)
- Type
- Slide-deck generation: plan or document in, editable PowerPoint deck out
- Architecture
- One LoRA adapter on a 20B base serving two roles: a designer that writes a JSON brief (palette, type system, per-slide layout, content and rationale) and a coder that writes PptxGenJS JavaScript, one call per slide
- Base model
- Not named in IBM's post
- Training data
- About 1,000 synthetic decks (around 10,000 slides) generated with a large open teacher model in IBM Design Language, including reasoning traces for design choices
- Directives
- Inline [[ ]] instructions from 12 categories, such as [[make this a timeline]]
- Harness
- Rule-based fixes, then a renderer and geometry checker for overlaps and overflow, then a separate editor model; a rewrite is kept only if the checker confirms it is better
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: because each slide's brief is fixed before coding, coder calls run in parallel and the coding stage is about 6–7x faster.
Lab claim: across 131 PresentBench decks, the check-and-repair harness adds 3.5 points on average.
Lab claim: on three PresentBench domains (IBM's runs; baselines from the PresentBench paper), Palette scores highest of the compared systems on Economics and Education, third on Talk, and second on the three-domain mean (59.7 vs NotebookLM's 60.8).
Sources
LABPalette: building slide decks in your design language (IBM Research, Hugging Face blog)huggingface.co/blog/ibm-research/palette-deck-building · Nov 3, 2026Tracking ends Feb 1, 2027 unless the API or weights become available.