Best Open Source Image Models 2026: FLUX 3 Klein and Qwen-Image-2.1 Ranking

👁 190 reads

The best open-weight image model at the end of October 2026 is a near tie: FLUX 3 Image Klein (1043 Elo) and Qwen-Image-2.1 (1042) top the open models on the Artificial Analysis leaderboard, with Bria’s new Fibo 2.0 three points behind. None of the three is free for every commercial use. If you need a permissive license, Ming-Image-0.1-Design (MIT, 999 Elo) is the highest-ranked option, and Qwen-Image-2512, Z-Image Turbo and FLUX.2 [klein] 4B are on Apache 2.0. Scores below were read on October 29, 2026.

Open weights are not the same as a free license

Most models here are open-weight, not open source in the strict sense. The difference is practical:

• Open weights means you can download the model file and run it locally, for example in ComfyUI.
• The license decides what you may do with the model and its images. Several top models forbid commercial use, and one gives no rights at all in the European Union.

I list the license for each model as stated on its Hugging Face card. Where I could not confirm it, the table says “check license on model page.”

FLUX 3 Image Klein and Qwen-Image-2.1 are one point apart

Artificial Analysis (AA) ranks models by blind votes: people see two images made from the same prompt and pick the better one. The score is in Elo points. The “Overall” column is the position among all models, closed ones included; the leader overall is GPT Image 2.5 Sunburst at 1222.

Bars start at 900 points. The full list of open-weight models on the leaderboard, with licenses:

A dash means the model has no score on that leaderboard. I read the Ideogram 4.0 license on the fp8 card in the official Ideogram 4 collection; the Fast and Instant variants may have different terms. In editing, Fibo 2.0 now leads the open models (1088), ahead of Qwen-Image-2.1 (1081) and FLUX 3 Image Klein (1072).

Two Apache 2.0 models sit outside this list. Qwen-Image-2512 (Alibaba, 20B) no longer appears on the leaderboard, and FLUX.2 [klein] 4B has an editing score (938) but no text-to-image score. Black Forest Labs says the 4B model fits in about 13 GB of VRAM, which makes it one of the few commercial options for 16 GB cards.

The top of the table is tight. The first three models are within 4 points, and each score carries a margin of about ±10, so treat them as one class and choose by license and hardware.

Licenses decide more than Elo points

• FLUX 3 Image Klein - Black Forest Labs says you may fine-tune and build on the weights within the license, but running Klein in a product “needs a commercial license.” Check the Hugging Face card before you ship anything.
• Qwen-Image-2.1 - the Qwen Research license allows research, evaluation and other non-commercial use. Commercial use needs a separate license from Alibaba. The license, clause by clause.
• Fibo 2.0 - the downloadable weights are non-commercial; Bria sells commercial rights separately.
• FLUX.2 [dev] and FLUX.2 [klein] 9B - FLUX Non-Commercial License. For commercial work Black Forest Labs offers its API.
• Ideogram 4.0 - “Ideogram 4 Non-Commercial.”
• HunyuanImage 3.0 Instruct - the Tencent Hunyuan Community License excludes the European Union, the United Kingdom and South Korea, so it grants nothing there. Elsewhere commercial use is allowed below 100 million monthly active users.
• Qwen-Image-2512, Z-Image Turbo, FLUX.2 [klein] 4B - Apache 2.0: commercial use allowed if you keep a copy of the license.
• Ming-Image-0.1-Design and HiDream-O1-Image - MIT: commercial use allowed if you keep the copyright notice.
• Cosmos3-Super-Text2Image - OpenMDW-1.1; NVIDIA’s card says the model is fit for commercial and non-commercial use.

The best fully permissive model, Ming-Image-0.1-Design (999), trails the leader by 44 points. If you sell graphics, that gap matters less than the legal risk of a non-commercial license.

FLUX 3 Image Klein peaked at 7.5 GB of VRAM, Qwen-Image-2.1 needs 24 GB

A local generator is only useful if it fits your GPU. I list only figures with a source, my own measurement or the model card:

On my RTX 4090, Qwen-Image-2.1 in int8 makes a 1024×1024 image in 8.4 seconds (12 steps) and a 2048×2048 image in 44 seconds. How to match the variant to a smaller card: Qwen-Image-2.1 VRAM requirements. My FLUX 3 Image Klein run is in FLUX 3 Klein local test.

For home hardware the cut is simple: FLUX 3 Image Klein, Z-Image Turbo and FLUX.2 [klein] 4B fit in 16 GB or less, Qwen-Image-2.1 wants 24 GB, and the 64-80B models need server GPUs even though their weights are open.

All three leaders run in ComfyUI

Qwen-Image-2.1 had ComfyUI support from day one, along with Diffusers, vLLM-Omni, SGLang and LightX2V. My ComfyUI installation guide for Windows includes a ready script. I ran FLUX 3 Image Klein in ComfyUI too, and Z-Image Turbo and the FLUX.2 family are covered in my Z-Image Turbo and FLUX tests.

Qwen-Image-2.1 still loses to closed models on text

The AA ranking measures which image people like more. It does not measure text rendering or layout precision. I ran Qwen-Image-2.1 through the same tests I use for closed models, with the on-image text in Polish, a language with diacritics that trips up most generators:

• Text rendering: 6 of 13 strings correct (the best closed models 13 of 13, Ideogram 4.5 11 of 13).
• Layout described in words: 2 of 16 objects in place, average IoU 0.17 (FLUX 3 with bounding boxes in JSON: 16 of 16).
• Small text in a large frame: 2 of 8 small signs readable at 2752×1536.

Details: Qwen-Image-2.1 vs closed models. If text on the image is your main use, closed models still win.

Which one to choose

Five steps to a first image:

1. Check your GPU memory (Windows: Task Manager, Performance tab, GPU).
2. Decide whether the images will earn money. If yes, FLUX 3 Image Klein (without a commercial license), Qwen-Image-2.1, Fibo 2.0, FLUX.2 [dev] and Ideogram 4.0 are out.
3. Pick from the table: up to 16 GB, Z-Image Turbo or FLUX.2 [klein] 4B for commercial work, FLUX 3 Image Klein for private work; at 24 GB, Qwen-Image-2.1 too.
4. Install ComfyUI and download the weights from Hugging Face.
5. Start at 1024×1024 and raise the resolution later: on my card, going from 1024 to 2048 pixels took Qwen from 8.4 to 44 seconds.

For a comparison with closed models, see the AI image generator ranking.

Frequently asked questions

Is there a truly open source image model?

Yes, in the licensing sense. Qwen-Image-2512, Z-Image Turbo and FLUX.2 [klein] 4B are on Apache 2.0, and Ming-Image-0.1-Design and HiDream-O1-Image on MIT. All allow commercial use if you keep the license or copyright notice.

Can I use FLUX.2 [dev] commercially?

No. FLUX.2 [dev] and FLUX.2 [klein] 9B are under the FLUX Non-Commercial License. Of the FLUX.2 family, only [klein] 4B is on Apache 2.0.

Can HunyuanImage 3.0 be used in the EU?

No. The Tencent Hunyuan Community License excludes the European Union, the United Kingdom and South Korea. The Instruct version also needs at least eight 80 GB cards.

See also: Local AI: text models, images and video on your own computer · Qwen-Image-2.1 on the Artificial Analysis ranking

// read moreFLUX 3 Image Klein: Local Test on an RTX 4090BenchmarksQwen-Image-2.1 Tops Open Models in Artificial Analysis RankingBenchmarksQwen-Image-2.1 Benchmark: Full Rankings of 29 Image ModelsBenchmarks

Kinetyra