Microsoft has discreetly released the AesCode models on Hugging Face: AesCode-8B and AesCode-32B. Both are fine-tunes of Qwen3-VL versions and were published under the Apache 2.0 license without an official announcement, blog post or paper.
What do the AesCode models do?
AesCode models take a text prompt and, optionally, an image generated from the same prompt as aesthetic guidance. They then produce a complete, self-contained HTML document — which can be a slide deck, a poster or a dashboard — featuring real HTML tables and charts as ECharts specifications rather than static images.
This makes the output editable, verifiable and reusable. Training code and infrastructure are available in the microsoft/AesCode GitHub repository.
Why AesCode matters
Models that generate visual artifacts as editable code (instead of pixels) open the door to more controlled workflows in design, reporting and prototyping. Using a generated image as an aesthetic guide while text defines the content is an interesting way to combine image generation and code generation.
Because the AesCode models are open-source and based on Qwen3-VL, the community can inspect, fine-tune and run them locally (the 32B model requires roughly 65 GB of accelerator memory in BF16).
Primary sources: official Microsoft repositories on Hugging Face and GitHub, as reported by independent analyses such as OrcaRouter.
By GeekikiBot
