Qwen-Image-2.1 Tops Open Models in Artificial Analysis Ranking

👁 137 reads

As of October 2, 2026, Qwen-Image-2.1 is the best open-weight model on both Artificial Analysis image leaderboards. In text-to-image it scores 1034 Elo points and ranks 18th overall, in image editing 1074 points and also 18th place. It beats FLUX.2 [dev], Ideogram 4.0 and HunyuanImage 3.0, though it still trails the closed leaders from OpenAI and Google by a significant margin.

Artificial Analysis (AA) announced the result in a post on X. The model launched on September 20, so it took just under two weeks for the full evaluation. We checked both leaderboards and compared the numbers with what we measured on our own card.

What Artificial Analysis Actually Did

AA runs rankings based on blind comparisons. Voters see two images from the same prompt, don’t know which model made them, and choose the better one. From thousands of votes, an Elo score is calculated. The reference point for the scale is FLUX.2 [dev] with a score of 1000.

There was one complication with Qwen-Image-2.1: on the test date, no major API provider offered this model. So AA hosted it themselves, on their own hardware, using the open weights. For this reason, the model shows a note “No API available” and has no price per 1000 images, which AA provides for models available through APIs.

The results in numbers:

• Text-to-image (AA-Image-T2I v2.0): 1034 points, ±10, from 5286 votes, 18th place overall, 1st among open models.
• Image editing (AA-Image-Editing v2.0): 1074 points, 95% confidence interval 1065-1083, 18th place overall, 1st among open models.

Bars are calculated from 900 points. The lead over second place is 23 points, significantly more than the ±10 margin of error. Behind Qwen, a whole group of models is clustered in the 995-1011 range, so there’s a tight race for second place among open models, but not for first.

In editing, second place goes to HunyuanImage 3.0 Instruct with 1066, third to the older Qwen Image Edit Plus 2511 with 1021.

Who Ranks Higher

To keep expectations realistic: 18th place means 17 models are rated better. In text-to-image, the leaders are:

• GPT Image 2.5 Sunburst (max) - 1197 points,
• GPT Image 2.5 Flare (max) - 1190,
• GPT Image 2 (high) - 1172,
• Grok Imagine Image 2.0 - 1155,
• MAI-Image-2.6 - 1150.

Between them and Qwen are paid models from the same company: Qwen-Image-3.0-Pro (1089) and Qwen-Image-3.0 (1075), available only through API. In editing, the situation is more interesting - the open 2.1 with 1074 beats regular Qwen-Image-3.0 (1055) and nearly matches the Pro version (1076).

Jump from Qwen Image 2.0

The previous version from March 2026 performed mediocrely in AA. The difference in six months:

• Text-to-image: from 72nd place (961 points) to 18th (1034).
• Editing: from 58th place to 18th.

Same producer, six months apart, and a completely different league. As a reminder: the generation component of Qwen-Image-2.1 has 7 billion parameters, one model handles both generation and editing, and images are created natively in 2K, including with transparent backgrounds.

LMArena Confirms, But on a Different Scale

AA also cited results from LMArena, the second popular voting-based ranking. The point scale there is different, so the numbers can’t be directly compared with AA:

• Image Edit Arena: 1367 points, 16th place overall, 1st among open models.
• Text-to-Image Arena: 1228 points, 17th place overall, 1st among open models.

Two independent rankings, two different sets of voters, and the same conclusion: as of today, there’s no better-rated open model for images.

What This Means in Practice

First, the good news. A model that ranks 18th in the world, we ran on a regular RTX 4090 card with 24 GB. The int8 variant weighs 16.6 GB, a 2048×2048 image takes 44 seconds to generate, and electricity for one 2K image costs us about $0.002. The closed models at the top are only available through APIs.

Now what the ranking doesn’t tell you. AA measures which image people consider better. It doesn’t measure whether text is error-free or whether objects are positioned where you specified. We tested this ourselves (we ran the writing tasks in Polish, a heavily inflected language with diacritics):

• Non-English text rendering: 6 out of 13 strings correct, while the best closed models get 13 out of 13, and Ideogram 4.5 gets 11 out of 13.
• Layout described in words: 2 out of 16 objects in their correct place, average IoU 0.17. FLUX 3 with bounding boxes in JSON hit 16 out of 16.
• Small text in 4K: 2 out of 8 small signs readable at maximum resolution of 2752×1536.

Voters like Qwen’s images, but if you’re creating graphics with text in your language or with precise composition, the ranking will mislead you. Full results are in the article Qwen-Image-2.1 vs. closed models.

It’s also worth remembering that at launch, Alibaba claimed the model beats Nano Banana 2. Independent AA measurement doesn’t confirm this: Nano Banana 2 has 1125 points in text-to-image and 1108 in editing.

The Catch: License

Open weights don’t mean an open license. Qwen-Image-2.1 is released under the Qwen Research License Agreement, which permits only non-commercial purposes, research, and evaluation. Commercial use requires a separate commercial license from Alibaba. This distinguishes it from some competitors on the same leaderboard - Qwen-Image-2512 is under Apache 2.0, and Ming-Image-0.1-Design under MIT. We compiled a comparison of licenses for all open models in our open image models ranking, and Qwen’s specific terms in a separate article.

In other words: if you want to generate images locally for learning, testing, and personal projects, this is currently the best choice. If you want to profit from them, this ranking changes little until you sign an agreement with Alibaba.

How to set up the model yourself, we described in the installation guide in ComfyUI on Windows, and how much memory you need on weaker cards - in the article about VRAM.

// read moreQwen-Image-2.1 Benchmark: Full Rankings of 29 Image ModelsBenchmarksBest Open Source Image Models 2026: FLUX 3 Klein and Qwen-Image-2.1 RankingBenchmarksQwen-Image-2.1 Alpha Channel: Native Transparent PNG TestBenchmarks

Kinetyra