Qwen-Image-2.1: a 7B open image model you can run but can't ship
Alibaba's 7B RGBA model runs on a 3090 and rivals closed generators, but its research-only license bars commercial use. Plus Willison's 2026-in-LLMs recap.
Copy markdown
A frontier open image model — with a license trap
Qwen-Image-2.1 is a 7B open-weight model with native RGBA transparency, support for up to 10 reference images (group shots, virtual try-ons, room design) and guided local edits by mask or scribble. Qwen says it beats most closed models on its own benchmark; independent numbers are still pending. The catch for builders: the weights ship under a research-only license that bars commercial use, so putting it in a product means applying to Qwen for a separate license.
Run it on a single 3090 this weekend
GGUF quants and ready-made ComfyUI workflows are already out, so the model runs locally on a 24GB card, with quantized builds squeezing onto 16GB+. Weekend build: a local transparent-PNG asset pipeline, a virtual try-on demo or room mockups with no per-image API bill — just keep the output non-commercial until you clear the license.
Willison's '2026 in LLMs' is the pre-DevDay catch-up
Simon Willison posted his closing WeAreDevelopers keynote, a chronological run through the year in LLMs, with the video plus a ~8,000-word annotated transcript. A fast way to reload context before OpenAI's DevDay keynote tomorrow at 10am PT.