Key takeaways
Qwen-Image 2.1 is the most capable open image model here, with generation, editing, transparent output and up to 10 references, but its license is research-only; Qwen-Image-2512 and Edit-2511 are the Apache 2.0 route for commercial work.
HiDream-O1-Image is an 8B, MIT-licensed model with native 2048 by 2048 output, instruction editing and layout control.
Z-Image Turbo runs on 16 GB consumer cards in eight steps, under Apache 2.0.
MOSS-Image, released in October 2026, places text and objects from a JSON layout.
Check the license first: two of the seven are non-commercial, and Tencent's excludes the EU, UK and South Korea.
How we ranked
Every model here has downloadable weights. Rank reflects output resolution and detail, editing and reference support, text rendering, the hardware needed to run it and how freely the license allows commercial use. Facts are from each official model card and license file, read November 2, 2026.
Open-weight image models at a glance
Model | Size | Max resolution | Editing | License |
|---|---|---|---|---|
Qwen-Image 2.1 | 7B generation component | 2048 x 2048 | Built in, with transparency | Research only (non-commercial) |
HiDream-O1-Image | 8B | 2048 x 2048 | Built in | MIT |
Z-Image / Turbo | 6B | 2048 x 2048 total area | Not yet released | Apache 2.0 |
MOSS-Image | 5B | - | Layout control only | Apache 2.0 |
LongCat-Image | 6B | - | Separate edit model | Apache 2.0 |
HunyuanImage 3.0-Instruct | 80B (13B active) | Set per prompt | Built in | Tencent Hunyuan Community (not EU, UK or South Korea) |
Fibo 2.0 | 8B | - | Separate edit model | Non-commercial |
1. Qwen-Image 2.1: Best for generation and editing in one model
Qwen-Image 2.1 is the latest open image model from Alibaba's Qwen team, released in September 2026.
What works: One model generates and edits, with native transparent (RGBA) output, editable transparent layers and subject extraction. It accepts up to 10 reference images and local edits marked with circles, painted annotations or masks, and renders at 2048 by 2048. A Turbo version from October cuts generation from 40 steps to eight.
What doesn't: The Qwen Research License limits it to research and evaluation, so commercial work needs a separate license. For client projects, Qwen-Image-2512 and Qwen-Image-Edit-2511 are Apache 2.0.
Best for: Experiments with layered, reference-heavy edits.
2. HiDream-O1-Image: Best MIT-licensed all-rounder
HiDream-O1-Image, from May 2026, skips the VAE and works directly on pixels.
What works: The 8B model needs no separate text encoder and generates natively up to 2048 by 2048. It handles instruction editing, subject personalization, long text rendering and storyboards, and a mid-May update added layout and skeleton conditioning. The MIT license allows commercial use.
What doesn't: Full precision needs a large GPU, so most home setups rely on community FP8 builds of about 10 GB.
Best for: Commercial work that needs editing and control in one permissive model.
3. Z-Image Turbo: Best for consumer GPUs
Z-Image comes from Alibaba's Tongyi-MAI team, with Turbo released in November 2025 and the base model in January 2026.
What works: The 6B Turbo model generates in eight steps and fits on 16 GB consumer cards, at any aspect ratio up to 2048 by 2048 total pixels. The base model supports fine-tuning, and both are Apache 2.0.
What doesn't: The announced editing model still hasn't been released.
Best for: Fast text-to-image on a single gaming GPU.
4. MOSS-Image: Best for layout control
MOSS-Image, a 5B model from OpenMOSS released in October 2026, takes JSON layouts as prompts.
What works: Prompts can be plain language, tag lists or JSON layouts that place text and objects in specific positions, which suits posters, packaging mockups and slides. Weights and code are Apache 2.0.
What doesn't: There's no instruction-based editing.
Best for: Designs where text and elements must land in exact spots.
5. LongCat-Image: Best for bilingual text
LongCat-Image is Meituan's bilingual text renderer, out since December 2025.
What works: The 6B model renders English and Chinese text, with character-level accuracy when the text is in quotation marks, and separate Edit and Edit-Turbo models handle changes. It needs about 17 GB with CPU offloading and is Apache 2.0.
What doesn't: Editing requires switching to a second model.
Best for: Signs, labels and graphics with Chinese and English text.
6. HunyuanImage 3.0-Instruct: Best for maximum quality on server hardware
Tencent's HunyuanImage 3.0 is an 80B mixture-of-experts model with 13B active parameters, and the Instruct version arrived in January 2026.
What works: Instruct handles generation, editing and multi-image fusion, with a reasoning step that expands prompts, and a distilled version runs in eight steps.
What doesn't: Instruct needs eight 80 GB data-center GPUs, and the license does not apply in the EU, UK or South Korea.
Best for: Studios with cloud GPUs outside those regions.
7. Fibo 2.0: Best for structured prompts
Fibo 2.0, Bria's 8B model from October 2026, splits a prompt into camera, lighting and composition fields.
What works: Structured JSON prompts give repeatable control over each part of an image, and a separate edit-turbo model handles changes.
What doesn't: The open weights are for non-commercial use, and commercial work requires an agreement with Bria.
Best for: Testing controllable generation before licensing it.
Kandinsky 5.0 Image Lite is an MIT-licensed alternative for text-to-image and image-to-image.
How to choose and run an open image model
Read the license first. MIT and Apache 2.0 allow commercial use; research and non-commercial licenses don't.
Match the model to your VRAM: Z-Image Turbo and FP8 builds of HiDream fit 16 GB cards, while HunyuanImage 3.0 needs a server.
Use Diffusers or each model's official repository to get started.
Put text you want rendered in quotation marks, which several models treat as literal text.
Use a dedicated edit model to change an existing image.
Frequently asked questions
What is the best open-weight image model?
Qwen-Image 2.1 is the most capable open-weight image model in November 2026, combining generation, editing and transparent output, but it's licensed for research only. HiDream-O1-Image is the best pick for commercial work under the MIT license.
Which open image models can I use commercially?
HiDream-O1-Image (MIT), Z-Image, MOSS-Image, LongCat-Image and Qwen-Image-2512 (all Apache 2.0) allow commercial use. Qwen-Image 2.1 and Fibo 2.0 are non-commercial, and HunyuanImage 3.0 excludes the EU, UK and South Korea.
Can I run an open image model on my own PC?
Yes. Z-Image Turbo fits a 16 GB graphics card, HiDream-O1-Image runs in FP8 builds of about 10 GB and LongCat-Image needs about 17 GB with offloading.
Which open image model is best for text in images?
MOSS-Image places text from JSON layouts, LongCat-Image renders English and Chinese text accurately and HiDream-O1-Image handles long text in the prompt.
Which open image models can edit photos?
Qwen-Image 2.1, HiDream-O1-Image and HunyuanImage 3.0-Instruct edit in the same model, while LongCat-Image, Qwen-Image and Fibo 2.0 offer separate edit models.