This week Microsoft dropped the AesCode series on HuggingFace (an open-source model hosting platform, often called the AI world's App Store) — 8B and 32B sizes — engineered to convert prompts into HTML/CSS code for PPTs, posters, and dashboards. What you get isn't an uneditable image, but a webpage where you can change fields and swap charts. This, we believe, is the critical distinction.
What this is
AesCode breaks "make a good-looking PPT" into two jobs: what to write, and how to lay it out. A code model produces HTML/CSS (structure, data, text) per the prompt, while a reference image handles only "does it look good." Two models run separately, then seven reward signals (training-time scoring signals) align them across layout hierarchy, color, text readability, and other dimensions. The base is Alibaba's Tongyi Qianwen Qwen3-VL-32B (Alibaba's open-source multimodal vision-language model); the end product is a directly editable webpage, not a dead PNG.
Industry view
Supporters argue AesCode hits the weak spot of commercial presentation tools like Gamma and Beautiful.ai — AI-generated images you can't tweak, forcing you to either regenerate or manually redraw; editable HTML lets enterprise IT plug into their own templates and compliance workflows.
But reservations are equally clear: the 32B model demands 24GB of VRAM (equivalent to an RTX 4090), so local deployment is no light lift; the open-source release still trails commercial-grade layout aesthetics, and benchmark scores (跑分) can't replace customer eye-appeal. Worth flagging explicitly: the foundation is Alibaba's Tongyi Qianwen — meaning a Chinese base model has been put on the open-source shelf by Microsoft itself.
Impact on regular people
For enterprise IT — AesCode is commercially usable open source, but the 32B model starts at 24GB single-card VRAM. Run the numbers on local deployment and ops costs.
For individual professionals — For internal training or client proposals, have AI produce an "editable draft" first, then polish manually — no more endless regeneration loops.
For the consumer market — Over the next 1–2 years, presentation tools will fork in two directions: the Canva school chasing mass-market aesthetics, and the AesCode line chasing customizability for the enterprise.