microsoft/AesCode-32B
AesCode-32B by Microsoft is a 33.4 billion parameter multimodal model based on Qwen3-VL-32B-Instruct, designed to generate structured, editable, and verifiable visual artifacts like slides, posters, and dashboards as HTML/CSS. It uniquely addresses the challenge of visual code generation by using an image reference for aesthetic guidance while adhering to content prompts, and employs graph-structured supervision and decoupled cross-modal rewards to separate semantic requirements from visual cues. This model excels at producing high-quality, editable visual code, leading benchmarks in overall performance and rule adherence for programmatic checks.
Loading preview...
AesCode-32B: Generating Editable Visual Artifacts as HTML/CSS
AesCode-32B is a 33.4 billion parameter multimodal model developed by Microsoft, built upon the Qwen3-VL-32B-Instruct architecture. Its primary function is to generate information-rich visual artifacts, such as slides, posters, and dashboards, directly as structured, editable, and verifiable HTML/CSS code. This model addresses a key challenge in code generation: enabling models to understand and apply visual layout, hierarchy, and color.
Key Capabilities
- Structured HTML/CSS Generation: Produces visual designs as editable code, ensuring verifiability and ease of modification.
- Multimodal Integration: Utilizes an image generated from the same prompt as an aesthetic reference, guiding visual composition while adhering to content requirements.
- Decoupled Cross-Modal Rewards: Employs graph-structured supervision and separate reward channels (execution, text, boundary, table/chart, layout, whitespace, design) to distinguish semantic content from visual cues, preventing hallucination and improving layout adherence.
- High Performance: Achieves an "Overall" score of 85.89% and a "Rule" score of 94.33% on infographic generation benchmarks, outperforming models like GPT-5.5 and Claude Opus 4.8 in programmatic checks.
- Robust Code Output: Generates HTML with proper table structures and ECharts specifications for charts, making the output directly inspectable and verifiable.
Good For
- Automated Visual Content Creation: Ideal for generating editable slides, posters, dashboards, and reports.
- Research in Multimodal Code Generation: Serves as a strong baseline and tool for exploring verifiable visual design and code synthesis.
- Developers Needing Structured UI Output: Useful for scenarios where visual designs need to be programmatically generated and easily modified or integrated into web applications.