Qwen-Image
Alibaba Qwen team's 20B MMDiT image foundation model plus its Edit and Layered variants, all Apache 2.0. Renders long-form Chinese and English text inside images at commercial quality.
Overview
Text inside generated images was the unsolved problem for years, and Chinese text was worse than unsolved. Qwen-Image, a 20B MMDiT released by Alibaba's Qwen team in August 2025 under Apache 2.0, is the model that fixed it for Hanzi. Multi-line paragraphs, correct stroke structure, signage, packaging copy: it holds together where models trained mostly on Latin script produce something that looks like writing from a distance and falls apart up close. That single capability is why it took over so much commercial design work in the mainland.
The family grew quickly. Qwen-Image-Edit brought mask-free instruction editing in August 2025, Edit-2509 added multi-image input, then Edit-2511 on 15 November 2025 attacked the drift problem where a subject slowly mutates across successive edits, and folded several popular community LoRAs straight into the base weights. Qwen-Image-Layered followed on 19 December 2025, splitting an output into separate layers instead of one flattened raster. Qwen-Image-2512 on 31 December 2025 is the current open flagship, better on human realism and fine natural detail than the original release.
Two practical notes before you plan around it. In bf16 the 20B backbone is roughly 40GB of weights, so an 80GB A100 or H100 is the comfortable target and the layer-by-layer offload path that squeezes into 4GB is slow enough to be a proof of concept. And the actual frontier is no longer downloadable: Qwen-Image-2.0 and 2.0 Pro launched on 10 February 2026 as API-only models on Alibaba Cloud Model Studio, with no open release announced as of late August 2026. Weights you can pin means 2512 and Edit-2511.
Key Features
- ✓ 20B MMDiT backbone released August 2025, refreshed as Qwen-Image-2512 on 31 December 2025 with better skin rendering and natural texture.
- ✓ Renders multi-line Chinese and English text inside images with correct stroke structure and layout, the capability the model was explicitly built around.
- ✓ Qwen-Image-Edit-2511, out 15 November 2025, does instruction-driven editing with no masking, fixes subject drift across successive edits and bakes several popular community LoRAs into the base weights.
- ✓ Qwen-Image-Layered, 19 December 2025, decomposes a generated image into separate layers for design handoff rather than flattening everything.
- ✓ A layer-by-layer CPU offload path brings inference down to roughly 4GB of VRAM at a heavy speed penalty.
- ✓ Native support in Diffusers, ComfyUI and DiffSynth-Studio, with weights on both Hugging Face and ModelScope.
Where it holds
- • Apache 2.0 on a 20B image model is unusually permissive at this parameter count, and it covers the Edit and Layered variants too.
- • Chinese text rendering is the standout capability: multi-line Hanzi lands correctly where models trained mostly on Latin script produce decorative noise.
- • Checkpoints are date-versioned (2509, 2511, 2512), so you can pin an exact build and know what you are shipping.
- • LoRA and full fine-tuning tooling is mature, with widely used community training scripts rather than a research-only reference implementation.
Where it breaks
- • Qwen-Image-2.0 and 2.0 Pro, launched 10 February 2026, are API-only with no announced open release. The open flagship is 2512, not the model Alibaba markets.
- • 20B in bf16 is around 40GB of weights. Comfortable single-GPU work means an 80GB A100 or H100, or aggressive quantization; the 4GB offload path is a demo, not a workflow.
- • Prompt behaviour is noticeably better in Chinese than English for typography-heavy work, which is a real gap if your team does not read Mandarin.
- • Version naming is opaque to newcomers. Nothing in the string 2511 tells you it supersedes 2509, and the docs assume you already know.
My Take
Type a paragraph of Chinese into almost any western image model and you get glyph soup. This one lays out multi-line Hanzi with correct stroke structure and sane spacing, which is why it displaced imported models across a lot of poster and e-commerce production here. 20B MMDiT, Apache 2.0, weights on Hugging Face and ModelScope, roughly 40GB in bf16 so budget an 80GB card for anything past experimentation. The catch is that Alibaba kept Qwen-Image-2.0 from February 2026 behind the API, so what you download is 2512 rather than the current flagship.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- All open checkpoints (Qwen-Image, Qwen-Image-2512, the Qwen-Image-Edit series and Qwen-Image-Layered) are free under Apache 2.0 with commercial use permitted and no registration. The closed Qwen-Image-2.0 and 2.0 Pro are metered per image through Alibaba Cloud Model Studio; the Singapore endpoint takes overseas accounts, the mainland console expects real-name verification.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Qwen-Image holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Qwen-Image
FLUX
paidImage generation and editing models from Black Forest Labs, available as a hosted API or as downloadable weights you run on your own GPUs.
Tongyi Wanxiang / 通义万相
freemiumAlibaba’s AI image generation platform offering text-to-image creation with strong aesthetic quality and Chinese cultural understanding.
Recraft
freemiumImage model built for design output rather than art, with true SVG vector generation, locked brand styles and an API covering vectorize, upscale and background removal.
Adobe Firefly
freemiumAdobe's creative AI workspace, now a router for 30-plus models including its own indemnified Firefly Image Model 5, wired into Photoshop, Express and an enterprise-only API.