IBM Palette: 20B Open Model Builds Decks in Your Design Language
By Ali Al-Montaser — Senior AI Technical Editor
IBM Research released Palette, a 20B open-weights model trained with LoRA on synthetic IBM Design Language data. It ranks first in Economics and Education on PresentBench and second overall with 59.7 points versus NotebookLM's 60.8. The model runs self-hosted, accelerates slide coding 6–7× via parallelism, and a repair harness adds 3.5 points across 131 decks.
Executive Overview
IBM Research has announced Palette via the official Hugging Face blog: a 20B open-weights model trained with LoRA on synthetic data that follows IBM Design Language, designed to build presentations in the user's own design language on their own infrastructure. Palette ranked first in Economics and Education on the PresentBench benchmark and second overall with an average of 59.7 points versus 60.8 for NotebookLM. A repair harness adds 3.5 points across 131 decks, and the slide coding stage runs 6–7× faster thanks to parallelism. The release is distributed through Hugging Face as a self-hosted model, keeping sensitive content inside enterprise systems.
📊 Official Data & Technical Specification Sheet
| Technical Axis | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Open-weights model, self-hosted; no official API pricing announced. Cost is purely operational on the user's infrastructure (20B-parameter compute + LoRA). |
| 🌐 Platforms & Immediate Availability | Officially published on the Hugging Face blog (ibm-research/palette-deck-building); distributed via Hugging Face. Runs on private infrastructure (self-hosted) to ensure content confidentiality. |
| ⚡ Performance & Speed Metrics | Coding stage 6–7× faster via slide-by-slide parallelism. Harness adds +3.5 points across 131 decks. PresentBench average: 59.7 points (second place). |
| 🛡️ Security & Breach Resistance | Core security advantage: sensitive content never leaves enterprise systems because the model runs locally. The harness prevents slide degradation (edits accepted only if the validator proves improvement). |
| 🧠 Context Window | Context window size not explicitly disclosed; however, the brief is fixed per slide before coding, enabling independent parallel calls per slide. |
| 🌍 Arabic & Regional Support | Not officially confirmed; training on IBM Design Language and English directives. No explicit Arabic support in the official statement. |
Deep-Dive Features & Architecture
Palette uses a two-role architecture served by a single LoRA adapter: a Designer role that reads the plan and writes a structured JSON brief including the color palette, font system, and for each slide its layout, exact content, and a short design rationale; and a Coder role that takes one slide's specification and writes pptxgenjs JavaScript code that renders it, preceded by its justification. This separation distributes two kinds of decisions: editorial judgment for the Designer and precise engineering for the Coder. Because each slide's context is fixed before any code is written, the Coder calls run in parallel, making the coding stage roughly 6–7× faster. The adapter was trained with LoRA rather than full fine-tuning, with a practical lesson: Coder samples are about 5× longer than Designer samples, which required resampling the Designer at a 5× factor to avoid undertraining it.
The synthetic data was generated by a large open teacher model: about 1,000 decks covering nearly 10,000 slides, with each example carrying a written justification as a counterfactual scenario that weighs an alternative and explains why it was rejected, so the model learns the rule, not just the choice. Every slide is rendered and checked for overlap and overflow by the same validator used at inference; broken slides are repaired and their justification rewritten. The faithfulness problem was solved with reference decks rather than rules, since the team found the teacher copies from references far more than it follows instructions.
Benchmark & Competitive Performance
PresentBench evaluates decks on five criteria: fundamentals, visual design, completeness, correctness, and faithfulness. Palette ranked first in Economics and Education, where structured content and numerical accuracy decide the outcome, and third in Talk, where closed systems excel at narrative decks. Across the three domains it ranked second with 59.7 points versus 60.8 for NotebookLM, a 1.1-point gap, ahead of Manus 1.6 and five other systems. The repair harness contributed +3.5 points on average across 131 decks, and the parallel coding stage delivered the 6–7× speedup.
Industry Impact & Enterprise Adoption
Palette's open-weights, self-hosted design is a direct response to enterprise confidentiality requirements: sensitive content never leaves the organization's systems because the model runs locally. The two-role architecture and the validator-driven repair harness make output quality auditable, while the LoRA adapter keeps deployment costs tied to a 20B-parameter model rather than a larger closed API. For teams that need presentations in a specific corporate design language, Palette offers a customizable alternative to closed systems such as NotebookLM, though Arabic-language support remains unconfirmed and requires independent testing.
Conclusion
IBM Research's Palette is a significant open-weights entry in AI presentation generation: a 20B model with a single LoRA adapter, a Designer/Coder split, a repair harness worth +3.5 points on 131 decks, and a 6–7× faster coding stage. Its 59.7 PresentBench average trails NotebookLM's 60.8 by 1.1 points but leads in Economics and Education, and its self-hosted distribution addresses enterprise security directly. The model is available now via Hugging Face at ibm-research/palette-deck-building.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Palette is a 20-billion-parameter (20B) open-weights language model developed by IBM Research to build presentations in the user's design language. It was trained using LoRA on synthetic data following IBM Design Language and runs on the user's own infrastructure.
Palette ranked first in Economics and Education, third in Talk, and second overall across the three domains with 59.7 points versus NotebookLM's 60.8 (a 1.1-point gap), ahead of Manus 1.6 and five other systems.
Because each slide's context is fixed in the brief before any code is written, the Coder calls run in parallel slide by slide, making the coding stage roughly 6–7× faster than sequential generation.
The harness checks each generated slide with an engineering validator that detects overlap, clipping, and overflow, then rewrites it with an editor model. Edits are accepted only if the validator confirms improvement. The harness added 3.5 points on average across 131 decks on PresentBench.
The official source does not mention explicit Arabic support; the model is trained on IBM Design Language and accepts English directives inside [[ ]] brackets across 12 categories. Arabic support is not officially confirmed and requires independent testing.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
