2026-09-21 · MondayPublished daily at 08:00
AI Daily
- 1
- 2StepFun releases flagship model Step 5 Preview, with weights to be open-sourced on October 15Official WeChat account: StepFun
- 3Qwen-Image-2.1 now supports ComfyUI, with open weights available for downloadTongyi Qianwen / Qwen
01Model Releases/Updates
Qwen open-sources Qwen-Image-2.1: unified 7B generation and editing with native support for transparent images
The Qwen team has open-sourced Qwen-Image-2.1, unifying text-to-image generation and image editing in a single model. Its visual generation component has just 7B parameters and natively supports generating and editing transparent images. The model supports up to 10 reference images, as well as local editing specified through circular, freehand, or separate masks. It improves inference efficiency through a mixed-granularity attention architecture and KV cache reuse, while also enhancing text rendering, portrait lighting, and fidelity for people and products. It covers tasks including panoramas, infographics, and storyboards.
Original textStepFun releases flagship model Step 5 Preview, with weights to be open-sourced on October 15
StepFun releases its flagship foundation model Step 5 Preview, using a sparse MoE architecture with 600B total parameters and 27B activated parameters. It supports a 1-million-token context window and text and visual inputs, scores 44 on the Artificial Analysis Intelligence Index, ranks among the top three open-source models worldwide, and costs 1/8 as much per task as Claude Opus 5.
Original text02Product Releases/Updates
Qwen-Image-2.1 now supports ComfyUI, with open weights available for download
Qwen has announced that Qwen-Image-2.1 now supports ComfyUI, with weights available. A single 7B checkpoint supports both image generation and editing, native 2K generation, instruction-based editing using up to 10 reference images at once, and RGBA output with an alpha channel.
Original text