The Newest Open-Weight Image Models and What Each Is Good At

Key takeaways

• Qwen-Image-2.1 is the best open-weight image model right now. It's the top open model on both Artificial Analysis and LMArena, and it generates and edits in one model with up to 10 reference images, but its license is research-only.

• The two top-ranked open models aren't licensed for paid work. Qwen-Image-2.1 and Bria's Fibo 2.0, the leading open model for editing, both need a separate commercial license.

• NVIDIA's Cosmos3-Super is the highest-ranked open model with a fully commercial license. It's a 64-billion-parameter model that needs a server of datacenter GPUs.

• The newest releases are small and specialized. MOSS-Image (October 2026) places text and images from a JSON layout, and Ming-Image-0.1-Design makes posters and UI with transparent backgrounds.

• The best open option for paid work on modest hardware is still a 2025 Qwen model. Qwen-Image-2512 and Qwen-Image-Edit-2511 are Apache-2.0 and have the biggest fine-tuning ecosystem.

Why ad teams look at open-weight image models

An open-weight model is one whose trained weights you can download and run on your own hardware. For a creative team, that buys three things a hosted generator can't: control over where the images and the product photos behind them live, fine-tuning on a brand's own products or style, and a cost per image that falls toward the price of the GPU time once volume is high.

The trade-offs are just as real. The best closed models still lead the public leaderboards by a wide margin, running a large model takes GPUs and an engineer, and "open" doesn't mean free to use in an ad. In 2026 the licenses have moved in two directions at once: a few makers have opened up with MIT and Apache-2.0 releases, while the makers with the highest-ranked open models have moved their newest weights to non-commercial terms. Check the license before anything below goes into paid media.

How we picked

We looked at open-weight image models released or updated between December 2025 and November 2026 and picked the ones with a clear strength for ad and product work. For each we checked the release date, the license and whether it allows commercial use, the rankings on the Artificial Analysis and LMArena text-to-image and image-editing leaderboards where the model appears, the hardware its maker states, and the tooling for running and fine-tuning it. Where a strength comes only from the maker's own benchmarks, we say so.

The models at a glance

1. Qwen-Image-2.1 - Best for top overall quality

Qwen-Image-2.1 is Alibaba's Qwen team's newest open image model, released September 20, 2026, with an eight-step Turbo version on October 9.

What works: It's the highest-ranked open-weight model on both public boards: first among open models on Artificial Analysis for text-to-image (an Elo of 1043), and 19th overall on LMArena's text-to-image board and 20th for editing, above every other open entry. One 7-billion-parameter model handles generation and editing. Qwen says it takes up to 10 reference images, preserves identity "for people and products," accepts edits marked with circles, scribbles, or masks, and outputs transparent PNGs at native 2K. Support in the diffusers library arrived on day one, along with LoRA training tools.

What doesn't: The new Qwen Research License allows research and evaluation only, so client work needs a separate commercial license from Alibaba. Every earlier Qwen-Image release was Apache-2.0, so don't assume the old terms carry over.

Best for: Teams testing what open models can do now, and those prepared to negotiate a commercial license.

2. Fibo 2.0 - Best for precise edits and licensed training data

Fibo 2.0 is Bria's 8-billion-parameter model, released October 18, 2026, with a faster Edit Turbo version the next day.

What works: It leads the open models on Artificial Analysis's image-editing board (1088) and sits third among open models for text-to-image. Generation and editing share one model, and prompts can be written as structured JSON that sets lighting, camera, composition, and style, which gives repeatable results across a batch. Masked and multi-reference edits are supported, such as dressing a person in a referenced garment. Bria says it was trained only on licensed data.

What doesn't: The weights are for non-commercial use, so paid work goes through Bria's commercial license, and Bria doesn't state the GPU it needs. It isn't ranked on LMArena.

Best for: Brands whose legal team cares about training data and teams that want structured, repeatable product edits.

3. MOSS-Image - Best for layout control with a small model

MOSS-Image is OpenMOSS's 5-billion-parameter model, released October 16, 2026, and the newest model here.

What works: You can describe a layout in JSON with bounding boxes, placing a headline, a product, and a logo exactly where an ad template needs them. The maker also points to strong text rendering, prompts of up to 4,000 tokens, and range from photorealism to illustration. It's Apache-2.0 for both weights and code, has fine-tuning support from launch, and a community trainer runs LoRA training with as little as 6 GB of VRAM.

What doesn't: It's too new to appear on either public leaderboard, so its quality claims are the maker's own, and early users have reported color banding in the fine-tuned version.

Best for: Teams that build templated ads and want a permissive model they can fine-tune cheaply.

4. Ming-Image-0.1-Design - Best for posters, UI, and infographics

Ming-Image-0.1-Design is inclusionAI's 6-billion-parameter design model, released September 17, 2026.

What works: It's built for text-heavy compositions: posters, UI screens, infographics, and complete designed layouts, with transparent backgrounds. A companion model, Design-Layer, splits an image into ordered, editable layers. On Artificial Analysis it scores 999 for text-to-image, ahead of the larger Cosmos3 and HunyuanImage models. It's MIT-licensed.

What doesn't: The maker's validated setup is one GPU with 80 GB of memory, a lot for a 6-billion-parameter model, and it's a specialist rather than a general photo model.

Best for: Teams producing promotional graphics, banners, and app-style creative where the type has to be right.

5. JoyAI-Image-Edit-Plus - Best for combining several reference images

JoyAI-Image-Edit-Plus is JD's editing model, released in June 2026.

What works: It takes several reference images plus an instruction and combines elements from each into one new image, such as a product, a setting, and a model from separate photos. The earlier JoyAI-Image-Edit adds spatial edits that suit product shots: moving or rotating an object and changing the camera angle. Both are Apache-2.0.

What doesn't: It isn't ranked on either leaderboard, its maker doesn't state the hardware it needs, and its diffusers pipeline hasn't been merged into the main library yet.

Best for: Teams composing product scenes from existing photos under a permissive license.

6. Cosmos3-Super - Best for a top-ranked, fully commercial model

Cosmos3-Super is NVIDIA's 64-billion-parameter text-to-image model, released May 31, 2026, with a four-step distilled version in July.

What works: It's the highest-ranked model on LMArena's text-to-image board that carries a fully permissive license: 27th with its agentic prompting mode and 34th without it. The OpenMDW-1.1 license places no restrictions on use or on the images it makes, and NVIDIA says the model is ready for commercial use. The distilled version runs in four steps without classifier-free guidance.

What doesn't: NVIDIA says it needs four to eight H100 or H200 GPUs, or a single B200, so it's a datacenter model. It's built for NVIDIA's physical-AI work rather than ad creative, and Artificial Analysis ranks it lower than LMArena does.

Best for: Companies with GPU infrastructure that need the strongest model they can use commercially without a separate deal.

7. HiDream-O1-Image - Best for long text and identity across scenes

HiDream-O1-Image is HiDream's 8-billion-parameter model, open-sourced May 8, 2026.

What works: One model handles text-to-image, editing, subject-driven personalization, and storyboards. HiDream highlights accurate text across several regions of an image and in several languages, keeping a person or product consistent across new scenes, and layout and pose conditioning. It works directly in pixels, with no separate VAE, and outputs up to 2048 x 2048. It's MIT-licensed.

What doesn't: It ranks 51st on LMArena, below the Qwen and Cosmos3 models, and there's no official diffusers support, so it runs from HiDream's own code.

Best for: Teams that want a permissive model for multi-shot campaigns with the same product or person.

8. ERNIE-Image - Best for text-heavy layouts on a 24 GB GPU

ERNIE-Image is Baidu's 8-billion-parameter model, released in April 2026 with an eight-step Turbo version.

What works: Baidu built it for dense, long-form, layout-sensitive text: posters, infographics, comics, storyboards, and multi-panel images, along with realistic photos. Baidu says it runs on consumer GPUs with 24 GB of memory, it has diffusers support, and it's Apache-2.0.

What doesn't: It sits mid-table on Artificial Analysis despite strong text scores in Baidu's own benchmarks, and it isn't on LMArena.

Best for: Teams with a single high-end workstation GPU that need readable text in generated layouts.

9. Qwen-Image-2512 and Qwen-Image-Edit-2511 - Best for commercial work with the biggest ecosystem

Qwen-Image-2512 and its editing sibling are the last Qwen image models under Apache-2.0, released in December 2025.

What works: They allow commercial use without a separate deal. Qwen-Image-2512 improved text rendering, including Chinese, and Edit-2511 cut image drift, improved character consistency, and improved industrial and product design generation. Qwen-Image-Layered, from the same release, splits an image into RGBA layers so each part can be edited alone. They have the deepest bench of community LoRAs and fast variants of any open image models, and Edit-2511 ranks 41st on LMArena's editing board.

What doesn't: At 20 billion parameters they're heavier than Qwen-Image-2.1 and score about 100 LMArena points lower.

Best for: Agencies and brands that want to fine-tune an open model on their products and use the results in paid ads.

Older models still worth knowing

• Z-Image-Turbo (Alibaba's Tongyi lab, November 2025): Apache-2.0, eight steps, fits in 16 GB of VRAM, and the most-downloaded model in this group. Its editing model still hasn't shipped.

• HunyuanImage-3.0-Instruct (Tencent, January 2026): ranks 30th on LMArena for editing, second among open models, but its license doesn't apply in the EU, UK, or South Korea, and it needs eight 80 GB GPUs.

Which one fits your team

• You want the best open quality and you're ready to negotiate a license: Qwen-Image-2.1.

• Your legal team asks about training data: Fibo 2.0.

• You build templated ads and want to fine-tune cheaply: MOSS-Image or Qwen-Image-2512.

• You make posters, banners, and UI-style creative: Ming-Image-0.1-Design or ERNIE-Image.

• You have datacenter GPUs and need a fully commercial license: Cosmos3-Super.

• You need one product or person across many scenes: HiDream-O1-Image or JoyAI-Image-Edit-Plus.

Frequently asked questions

What is the best open-source image model in 2026?

Qwen-Image-2.1 ranks highest among open-weight models on both Artificial Analysis and LMArena. It's licensed for research only, though, so for paid work the strongest options are Cosmos3-Super if you have datacenter GPUs, or Qwen-Image-2512 and MOSS-Image on smaller hardware.

Can I use open-weight image models in commercial ads?

Some of them. MOSS-Image, JoyAI-Image-Edit-Plus, ERNIE-Image, and Qwen-Image-2512 are Apache-2.0, Ming-Image and HiDream-O1-Image are MIT, and Cosmos3-Super uses OpenMDW-1.1, all of which allow commercial use. Qwen-Image-2.1 and Fibo 2.0 need a separate commercial license, and HunyuanImage 3.0 isn't licensed in the EU, UK, or South Korea.

What GPU do I need to run an open image model?

It ranges widely. Z-Image-Turbo fits in 16 GB of VRAM and ERNIE-Image runs on a 24 GB card. Ming-Image's validated setup is one 80 GB GPU, and Cosmos3-Super and HunyuanImage 3.0 need a server with several 80 GB GPUs.

Are open-weight image models as good as closed ones?

Not yet at the top. The leading closed models score well above the best open model on both leaderboards, and Qwen-Image-2.1, the top open entry, ranks 19th on LMArena's text-to-image board. The open models are competitive for specific jobs, such as layout control, text-heavy design, and fine-tuning on your own products.

Which open image model is best for product photos?

For edits that keep a real product intact, Fibo 2.0 and Qwen-Image-2.1 rank highest, and Qwen says 2.1 preserves product identity across references. For a permissive license, Qwen-Image-Edit-2511 and JoyAI-Image-Edit-Plus can combine a product with new scenes. Check the label and shape in every output before it runs.

Do open-weight models make transparent PNGs?

Several do. Qwen-Image-2.1 and Ming-Image-0.1-Design output transparent images directly, and Qwen-Image-Layered splits an existing image into transparent layers.