Microsoft Just Dropped a Free AI That Builds Websites From a Sentence
Microsoft released two open-weights AI models that turn a plain-English prompt into a complete, editable HTML document — a landing page, a slide deck, a poster, or a full analytics dashboard. The project is called AesCode, the weights are free under the Apache 2.0 license, and the full training code is on GitHub. There was no press release, no keynote, no blog post. The models simply appeared.
That quiet arrival is part of the story. While the biggest labs fight over closed flagship models, Microsoft just handed everyone the weights to a system that, on its own benchmark, outperforms GPT-5.5 and Claude Opus 4.8 at generating visual documents as code.
What AesCode actually does
Most AI website builders give you one of two things: pretty pictures with garbled text, or clean code that looks like it was designed in 2004. AesCode attacks the gap between the two.
Here is the idea, straight from Microsoft’s research team: an image generator creates a reference layout from your prompt, and the code model uses that image purely as aesthetic guidance — for layout, hierarchy, and style. The prompt itself stays the source of truth for the actual content, so the model cannot copy the image generator’s hallucinated text or fake numbers. The output is a complete HTML document with real HTML tables and ECharts chart specifications, which means every number on the page is inspectable and verifiable.
The training behind it is unusually rigorous. Microsoft represents each target page as a “design graph” — nodes for text, charts, tables, cards, and regions, with edges for alignment and containment — and scores outputs across seven separate reward channels: execution, exact text, boundaries, table/chart data, layout, whitespace, and design. A deterministic verifier checks everything a program can check; a vision model judges the rest.
The numbers that matter
Microsoft evaluated AesCode on 300 infographic samples. The headline results:
- AesCode-32B scored 85.89 overall, ahead of GPT-5.5 (81.28) and Claude Opus 4.8 (80.39) given the same reference image.
- Exact text accuracy hit 95.34% — the model reproduces required copy nearly perfectly.
- Table and chart data accuracy reached 90.37%, the best in the comparison. If your dashboard says revenue grew 34%, that number is actually in the data, not painted on.
- Boundary failures dropped to 4.3%, versus 34.7% for GPT-5.5. In plain terms: elements stop overflowing the canvas.
Two checkpoints are available: AesCode-32B, the flagship, and AesCode-8B, a smaller version fine-tuned from Qwen3-VL-8B that can run on serious consumer hardware. Notably, the reference image is optional at inference time — withholding it costs the 8B model only about one point of visual quality, which means the design knowledge is baked into the weights, not dependent on the trick.
The money angle: client deliverables in minutes, zero API bills
This is where AesCode gets interesting for anyone who sells creative work. A freelancer who charges for pitch decks, one-page sites, dashboards, or marketing posters can now generate a structured, editable first draft from a prompt — then refine it by hand instead of starting from a blank canvas. The output is real HTML and CSS, not a flattened image, so you can edit text, swap colors, and fix charts directly.
Because the weights are open under Apache 2.0, there is no per-generation API fee and no vendor lock-in. Run the 8B model locally and your marginal cost per deck is electricity. Agencies producing high volumes of similar documents — real-estate flyers, restaurant menus, SaaS landing pages, investor updates — can build the model into an internal pipeline and keep the margin.
How to try it today
The models are on Hugging Face as microsoft/AesCode-32B and microsoft/AesCode-8B, and the training code, verifier stack, and an inference pipeline live at github.com/microsoft/AesCode. The 32B model needs roughly 65 GB of accelerator memory for the weights alone, so most people will want to serve it with vLLM or start with the 8B version. Microsoft includes a one-line pipeline that turns a short idea into the prompt-and-reference pair the model expects, plus a complete hospital-dashboard example to copy.
One honest caveat from Microsoft’s own report: “style” — whether a design needs zero revision before delivery — is the shared ceiling for every system tested, and no model clears 60%. Treat AesCode as the fastest junior designer you have ever hired, not as a replacement for taste. The draft is free and instant; the final 10% is still yours.
The bottom line
Open, free, and measurably better than the closed giants at this specific job: AesCode is the kind of release that quietly rearranges who can afford to produce professional visual work. If you make decks, dashboards, or pages for money, this one is worth a weekend.
Sources: AesCode GitHub repository (Microsoft, official); microsoft/AesCode-32B model card (Hugging Face, official).