Microsoft Open-Sources AesCode 8B/32B for Structured, Editable HTML Visual Artifacts
jacek2023 · reddit · 2026-10-11
Microsoft has released open weights for AesCode 8B and 32B, models that generate information-rich visual artifacts (slides, posters, dashboards) as structured, editable HTML/CSS.
Key details:
- Code models can't see how layout, hierarchy, and color land on the canvas; image generators look good but misrender text and numbers. AesCode uses an image generated from the same prompt as an aesthetic reference while following the prompt for content.
- Reference input alone isn't enough—off-the-shelf VLMs copy hallucinated content or ignore layouts—so AesCode uses graph-structured supervision and decoupled cross-modal rewards.
- AesCode-32B is built on Qwen3-VL-32B-Instruct with cold-start SFT followed by GDPO across seven reward channels, leading aggregate benchmarks.
Weights are on Hugging Face, including GGUF quants from bartowski.
More from coding & agent
- RL post-training Qwen 27B as a Cypher agent lifts graph-query accuracy 6.2 points for $119 — sophiamyang · 2026-10-11
- Same Qwen3-Coder 30B: instant success via LM Studio, 57-minute fix loop via Ollama — Proof_Nothing_7711 · 2026-10-11
- Insomnia keeps Mac agents running with the lid closed — engineers warn it overheats — emax · 2026-10-11
- Why most companies shouldn't be using AI agents yet: hype wastes millions daily — DavidLinthicum · 2026-10-11
- Typedef workshop: building a context graph layer for data and coding agents — AI Engineer · 2026-10-11
- Legwork: an MCP server that lets your AI find and sandbox-install GitHub tools — Status_Record5772 · 2026-10-11