Qwen-Image
Alibaba (QwenLM)
Qwen-Image is an image generation model line from Alibaba (QwenLM). The latest release is Qwen-Image-2.1 (20 Sept 2026). This page tracks all 6 Qwen-Image releases since 4 Aug 2025, with specs, licensing, and how to access each one. Official site →
Qwen-Image-2.1
Open-weight 7B image generation and editing model with native RGBA transparency and up to 10 reference images in a single pass; 2K generation
- License
- Qwen Research License
- Commercial use
- Restricted
- Availability
- Community API, Self-host
- Line
- Open-weight line, parallel to the closed hosted 3.0 - not a successor to it
- Parameters
- 7B (visual generation component)
- Architecture
- 32 single-stream DiT layers
- Capabilities
- image generation, editing, native transparency (RGBA), multi-reference (up to 10 images), subject extraction
Open-weight release that lands AFTER the higher-numbered 3.0: the two are parallel lines, not a sequence. 3.0 (2026-07-21) is the hosted, closed flagship - no weights, no model card, API only - while 2.x continues as the downloadable line. A lower version number here means newer, not older. Research license only; commercial use requires a separate Qwen commercial licence from Alibaba. Runs on consumer-grade GPUs. Claims benchmark performance competitive with larger closed models.
Qwen-Image-3.0
Prompts up to ~4,500 tokens (~4.5x 2.0) used for single-pass dense layouts - newspaper pages, exam papers with maths notation, multi-panel comics, infographic grids; legible text down to ~10px; 12 languages; unified generation and editing
- License
- Proprietary
- Commercial use
- Allowed
- Availability
- Official API, Consumer app
Break from the family's open-weights history: announced with no weights, license, parameter count, benchmarks or technical report - hence weights: closed despite qwen-image being a self-host family. Invite-only API on Alibaba Cloud Bailian / the Qwen platform from 2026-07-21; opened to all users 2026-08-05. No standalone announcement page located (the QwenLM GitHub blog feed stops at Qwen-Image-Edit), so announcement_url is null.
Qwen-Image-2.0
Next-generation model: smaller architecture with faster inference; 1k-token instruction support for infographics and posters; native 2K resolution; unified generation and editing in a single model
- License
- Proprietary
- Commercial use
- Allowed
- Availability
- Official API, Consumer app
No public checkpoint released as of launch - Qwen Chat and DashScope API only. Prior releases in this family (1.0, 2512, Edit) have open weights on HuggingFace.
Qwen-Image-2512
December upgrade: dramatically reduced AI look in human generation; finer natural textures (landscapes, fur, water); stronger text layout and text-image composition; #1 open-source on Alibaba AI Arena at launch
- License
- Apache 2.0
- Commercial use
- Allowed
- Availability
- Self-host, Community API
Named for its release month (December 2512 = 2025-12). Release date confirmed from README news entry.
Qwen-Image-Edit
Image editing model built on 20B Qwen-Image: text, semantic, and appearance editing; refreshes 2509 and 2511 followed
- License
- Apache 2.0
- Commercial use
- Allowed
- Availability
- Self-host, Community API
Qwen-Image
QwenLM team's dedicated image generation model (distinct from Qwen-VL); strong complex text rendering including Chinese; competitive with closed-source models on AI Arena at launch
- License
- Apache 2.0
- Commercial use
- Allowed
- Availability
- Self-host, Community API
- Parameters
- 20B
- Architecture
- MMDiT
Separate product from Qwen-VL (multimodal LLM) and Tongyi Wanxiang - different team and architecture. Release date confirmed from README news entry.
See Qwen-Image in context on the image model timeline, or browse all model families.