What is Palette and why do small models struggle with slides?

Frontier artificial intelligence models can create slide decks, but they typically share an identical house style and require sending confidential enterprise content to external servers. Qwen 3.8 27B en mi portátil demuestra que los modelos locales son útiles, pero las arquitecturas pequeñas suelen fallar al generar diapositivas, provocando títulos superpuestos o texto desbordado. Palette resuelve este inconveniente ajustando un modelo de 20 mil millones de parámetros con pesos abiertos.

El sistema entrena el modelo con datos sintéticos basados en el lenguaje de diseño de IBM Sebastian Raschka Lanza Curso Práctico sobre Modelos de Razonamiento, ayudando a mantener la coherencia tipográfica y visual. De este modo, las empresas pueden mantener la propiedad total de sus resultados ejecutando los modelos en su propia infraestructura.

How does the designer and coder architecture work?

Palette divides the generation pipeline into two distinct roles handled by a single LoRA adapter. The designer reads the high-level plan and outputs a structured JSON brief covering the palette, typography, layout, content, and design rationale.

The coder takes each individual slide specification from the brief and generates the necessary pptxgenjs code alongside its internal reasoning. By separating editorial judgment from precise geometry, the coder can execute tasks in parallel for each slide, speeding up the generation phase significantly.

How are synthetic datasets and validation harnesses built?

Generating quality decks without a public corpus requires a large open teacher model guided by detailed recipes and counterfactual reasoning traces. Every synthesized deck is evaluated by an automated checker, and any flaws found are corrected and fed back into the training loop as improved examples.

To ensure absolute reliability, Palette passes every generated deck through a check-and-repair harness before returning it. StateBench: Evaluando la Persistencia de Estado ayuda a comprender cómo los modelos procesan contextos largos, un principio similar al que utiliza este arnés para detectar solapamientos y corregir errores visuales.