GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All Posts
Qwen Image 2 vs 3: Which Is Newer, Open, and Better
2026/10/07

Qwen Image 2 vs 3: Which Is Newer, Open, and Better

Qwen Image 2 vs 3, plus 2.1: release dates, which versions have open weights, what changed in prompts, text and editing, arena scores, and which model to call.

Qwen Image 2 vs 3 sounds like a simple old-versus-new question, but the version numbers do not follow release order. Qwen-Image-3.0 came out on July 21, 2026[1]. Qwen-Image-2.1 came out two months later, on September 20, 2026, and it is the one with open weights[2]. That is why people on Reddit ask whether 2.1 is newer than 3.0.

This article lays out the whole lineup with dates, explains what changed between 2.0, 3.0 and 2.1, shows how they compare in blind human votes, and ends with which model to call for which job.

TL;DR

  • Release order: Qwen-Image-2.0 on February 10, 2026[3], Qwen-Image-3.0 on July 21, 2026[1], Qwen-Image-2.1 on September 20, 2026[2].
  • Open weights: of these three, only 2.1 has published weights, under the Qwen Research License[4]. The first Qwen-Image from August 2025 was released under Apache 2.0[5].
  • Biggest change from 2 to 3: prompts went from 1k tokens to 4.5k tokens, and Qwen says text renders legibly down to 10px in 12 languages[3][1].
  • Quality: on Artificial Analysis, Qwen-Image-3.0-Pro scores 1088 and Qwen Image 2.0 (2026-03-03) scores 963 for text-to-image[6].
  • 2.1 is not a 3.0 upgrade: it scores lower than 3.0 for text-to-image, but matches 3.0 Pro in editing and adds transparency, masks and up to 10 references[6][7][2].

The Qwen Image release timeline

ReleaseDateWeightsHeadline change
Qwen-ImageAug 4, 2025Yes, Apache 2.0[5]Precise text rendering[3], with a technical report[5]
Qwen-Image-2512Dec 2025Yes, Apache 2.0[8]Detail and photorealism[3]
Qwen-Image-2.0Feb 10, 2026None published[8]1k-token prompts, native 2K, one model for generation and editing[3]
Qwen-Image-3.0Jul 21, 2026None published[8]4.5k-token prompts, 10px text, 12 languages[1]
Qwen-Image-2.1Sep 20, 2026Yes, Qwen Research License[4]7B visual generator, transparency, up to 10 references[2]

"None published" means that on October 7, 2026, the Qwen organization on Hugging Face had no Qwen-Image-2.0 or Qwen-Image-3.0 repository. Its image repositories were Qwen-Image, the Qwen-Image-Edit releases, Qwen-Image-Layered, Qwen-Image-2512 and the Qwen-Image-2.1 releases[8].

The 2.0 post explains how the line got here. Before 2.0, Qwen ran two tracks: generation (Qwen-Image, then Qwen-Image-2512) and editing (single-image editing in August 2025, multi-image editing in September, consistency in December). Qwen-Image-2.0 merged them into one model[3]. Qwen-Image-3.0 is the third generation of that line and still handles both generation and editing[9]. Qwen-Image-2.1 is a smaller model released openly next to it.

Qwen Image 2 vs 3: what actually changed

Qwen sums up each generation with keywords. For 2.0 it listed "Precision, Variety, Completeness, Beauty, and Authenticity". For 3.0 it chose one word, "Real"[1]. In practice that means three changes:

  1. Longer prompts. Qwen-Image-2.0 supports 1k-token instructions[3]. Qwen-Image-3.0 supports up to 4.5k tokens. Qwen's example is a 3×3 grid of nine separate infographics described in 3.7k tokens and generated in one pass[1].
  2. Smaller text. Qwen says 3.0 renders text "as small as 10px" and can reproduce a full page of an academic paper with LaTeX formulas[1].
  3. More knowledge. 3.0 renders 12 languages natively, imitates web, game and livestream interfaces, and can look up current information online, such as a weather forecast for a given day[1].

Resolution did not change much. Qwen-Image-2.0 introduced native 2K[3], and Alibaba's API caps the 3.0 series at 2048×2048 total pixels[9].

The blind votes show the Qwen Image 2 vs 3 jump. Artificial Analysis ranks models by Elo from pairwise human preference votes[6]:

Model (as listed by Artificial Analysis)Text-to-image EloEditing Elo
Qwen-Image-3.0-Pro10881077
Qwen-Image-3.010751056
Qwen-Image-2.110351074
Qwen Image 2.0 Pro (2026-04-22)10321021
Qwen Image 2.0 (2026-03-03)963965

Text-to-image scores are from the text-to-image board[6] and editing scores from the editing board[7], both retrieved October 7, 2026. Both 3.0 tiers score above both 2.0 snapshots on each board.

Why Qwen-Image-2.1 is newer but not a 3.0 replacement

Qwen-Image-2.1 is a different kind of release. Qwen describes it as a model that "balances generation quality, inference efficiency, and cost", with a 7B-parameter visual generation component built from 32 single-stream DiT layers[2]. It is the open, compact branch of the family, not the next flagship.

It also adds things 3.0 does not have. Qwen-Image-2.1 can output transparent images, accepts up to 10 reference images, and takes local edits marked with circles, painted annotations or a separate mask[2]. Alibaba's API gives the 3.0 series 1 to 3 input images[9].

The votes match that split. For text-to-image, 2.1 (1035) trails both 3.0 tiers[6]. In the layout category the gap is wider: 1085 for 2.1, 1160 for Qwen-Image-3.0 and 1181 for Qwen-Image-3.0-Pro[10]. For editing, 2.1 scores 1074, close to 3.0 Pro at 1077 and ahead of standard 3.0 at 1056[7].

Our reading: use 3.0 when the image has to carry dense text or a complex layout. Use 2.1 for editing work, especially with many references, cutouts or transparent assets. If you want to run a model on your own GPU, 2.1 is the newest one you can download; Is Qwen Image 3 open source? covers that question in full.

Which Qwen Image model to call on reAPI

All three are on reAPI behind the same /api/v1/images/generations endpoint, with different request shapes:

Qwen Image 2Qwen Image 2.1Qwen Image 3.0
Model idqwen-image-2qwen-image-2.1qwen-image-3.0, qwen-image-3.0-pro
Prompt limit800 characters[11]5,000 characters[12]About 4.5k tokens[13]
Reference imagesOne source image[11]1 to 10, plus an optional mask[12]1 to 3[13]
Images per requestOne[11]One[12]1 to 6[13]
ResolutionUp to 2K[11]1K or 2K[12]1K or 2K[13]
Transparent outputNoYes, PNG or WebP[12]No
BillingFlat per image[11]Per image, 2K costs twice 1K, references free[12]Per image; Standard costs the same at 1K and 2K, Pro charges twice for 2K, references billed once per request[13]

Here is how we would settle Qwen Image 2 vs 3 for each job:

  • Posters, menus, newspapers, slides and infographics: Qwen Image 3.0. Use qwen-image-3.0-pro when the page carries a lot of small text.
  • Product edits, multi-reference composites, transparent assets: Qwen Image 2.1.
  • Already calling Qwen Image 2: move to 3.0 when you need longer prompts, more outputs per request or the higher scores above.

Current rates are on the Qwen Image 3.0, Qwen Image 2.1 and Qwen Image 2 model pages. For the cost of each 3.0 tier, see Qwen Image 3.0 API pricing.

FAQ

Is Qwen image 2.1 newer than Qwen image 3.0?

Yes, by release date. Qwen-Image-2.1 was released on September 20, 2026[2], two months after Qwen-Image-3.0 on July 21, 2026[1]. It is a compact open-weight model, not a successor: 3.0 still scores higher for text-to-image on Artificial Analysis[6].

Qwen image 3 release date

Qwen published the Qwen-Image-3.0 announcement on July 21, 2026[1]. Alibaba's API offers it as qwen-image-3.0 and qwen-image-3.0-pro[9].

Is Qwen Image 2.0 open source?

No weights have been published. On October 7, 2026, the Qwen organization on Hugging Face had no Qwen-Image-2.0 repository[8]. The open releases closest to it are Qwen-Image-2512 from December 2025 (Apache 2.0) and Qwen-Image-2.1 from September 2026 (Qwen Research License)[8][4].

Which Qwen models are open-source?

In the image line on Hugging Face: Qwen-Image, Qwen-Image-Edit and its 2509 and 2511 updates, Qwen-Image-Layered and Qwen-Image-2512 under Apache 2.0, and Qwen-Image-2.1 under the Qwen Research License[8][4]. Qwen-Image-2.0 and Qwen-Image-3.0 have no published weights.

Can you run Qwen image locally?

Only the releases with published weights, such as Qwen-Image or Qwen-Image-2.1[8]. Qwen Image 3.0 is available through APIs only; the details are in Is Qwen Image 3 open source?.

Is the Qwen image any good?

In blind human votes on Artificial Analysis, Qwen-Image-3.0-Pro ranks 14th on the default text-to-image view with 1088 Elo, behind GPT Image 2 (high) at 1172 and Nano Banana 2 at 1126[6]. In the layout category it ranks sixth[10]. We compare it with GPT Image 2 in Qwen Image 3 vs GPT Image 2.

Is the Qwen image free to use?

The open-weight Qwen-Image release is free to download under Apache 2.0[5]. Qwen Image 3.0 is a paid API. In Alibaba's Singapore region, Qwen-Image-3.0 comes with a free quota of 10 images, valid for 90 days[14]. Per-image costs are covered in Qwen Image 3.0 API pricing.

Picking a Qwen Image version today

The version numbers are misleading, but the choice is not hard. Qwen-Image-3.0 is the newest flagship, with the longest prompts and the best scores in the family. Qwen-Image-2.1 is newer by date, smaller and open, and it is the better editor for transparent assets and many references. Qwen-Image-2.0 is the earlier model both build on. For most new work on Qwen Image 2 vs 3, start with 3.0 and keep 2.1 for edits. Both can be tried from the Qwen Image 3.0 and Qwen Image 2.1 model pages.

References

  1. Qwen Team. Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge. Retrieved October 2026 from qwen.ai/blog?id=qwen-image-3.0
  2. Qwen Team. Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation. Retrieved October 2026 from qwen.ai/blog?id=qwen-image-2.1
  3. Qwen Team. Qwen-Image-2.0: Professional infographics, exquisite photorealism. Retrieved October 2026 from qwen.ai/blog?id=qwen-image-2.0
  4. Qwen. Qwen-Image-2.1 model card. Retrieved October 2026 from huggingface.co/Qwen/Qwen-Image-2.1
  5. Qwen. Qwen-Image model card. Retrieved October 2026 from huggingface.co/Qwen/Qwen-Image
  6. Artificial Analysis. Text to Image leaderboard (AA-Image-T2I v2.0). Retrieved October 2026 from artificialanalysis.ai/image/leaderboard/text-to-image
  7. Artificial Analysis. Image Editing leaderboard (AA-Image-Editing v2.0). Retrieved October 2026 from artificialanalysis.ai/image/leaderboard/editing
  8. Hugging Face. Qwen organization model list (search: Qwen-Image). Retrieved October 2026 from huggingface.co/Qwen
  9. Alibaba Cloud. Qwen Image Generation and Editing API reference. Retrieved October 2026 from alibabacloud.com/help/en/model-studio/qwen-image-generation-and-editing-api-reference
  10. Artificial Analysis. Text to Image leaderboard, Layout category. Retrieved October 2026 from artificialanalysis.ai/image/leaderboard/text-to-image?category=layout
  11. reAPI. Qwen Image 2 API reference. Retrieved October 2026 from reapi.ai/docs/qwen-image-2
  12. reAPI. Qwen Image 2.1 API reference. Retrieved October 2026 from reapi.ai/docs/qwen-image-2-1
  13. reAPI. Qwen Image 3.0 API reference. Retrieved October 2026 from reapi.ai/docs/qwen-image-3
  14. Alibaba Cloud. Model Studio: model inference pricing. Retrieved October 2026 from alibabacloud.com/help/en/model-studio/model-pricing
All Posts

More Posts