What is Qwen-Image-2.1?

Qwen-Image-2.1 is an open-weight text-to-image and image-editing model released by Alibaba's Qwen team on September 20, 2026 on Hugging Face, ModelScope and GitHub. Its visual generation component has 7B parameters across 32 single-stream DiT layers, paired with a Qwen3-VL 8B text encoder and an RGBA-capable VAE, and it generates natively at 2048x2048 in 40 steps. It is the first mainstream open model to generate and edit transparent (RGBA) images without a separate matting step, accepts up to 10 reference images for group portraits, try-on and room design, and supports local edits via circles, painted annotations or masks. It ships under the Qwen Research License, which bars commercial use without a separate grant from Alibaba.

Is Qwen-Image-2.1 better than FLUX.2 [dev]?

On Alibaba's own benchmark Qwen-Image-2.1 claims to beat most closed models, but as of September 21, 2026 there are no independent results, so treat quality claims as vendor numbers. What is objectively different: Qwen-Image-2.1 is 7B and runs on a 24GB consumer GPU such as an RTX 3090, generates transparent RGBA images natively, and takes 10 references; FLUX.2 [dev] is 32B, needs an fp8 build to fit consumer cards, has no native alpha channel, also takes up to 10 references, and edits at up to 4 megapixels. Licensing favours FLUX for business use: FLUX.2 [dev] uses BFL's non-commercial licence with paid commercial options through BFL, while Qwen-Image-2.1 requires applying to Alibaba for any commercial licence.

Can I use Qwen-Image-2.1 commercially?

Not under the default licence. Qwen-Image-2.1 is released under the Qwen Research License Agreement, which permits non-commercial and research use only; commercial deployment requires a separate licence from Alibaba and products must display 'Built with Qwen'. This is a step back from Qwen-Image 1.0 (August 2025), which was Apache-2.0. Teams needing a permissive open image model should look at Z-Image-Turbo (Apache-2.0, 6B) or FLUX.2 [klein] instead; teams that just need an API should use GPT Image 2.5 or Nano Banana Pro. As of October 2026, BFL's FLUX 3 Image Klein (5.6B open weights, FLUX Kommunity License) is another small open option to check against your use case.

Which image model is best for transparent PNGs in 2026?

Two options as of September 2026. GPT Image 2.5 (OpenAI, September 8, 2026) generates transparent backgrounds natively through the Images API and in ChatGPT and is the best hosted choice. Qwen-Image-2.1 is the only open-weight model with a true RGBA pipeline: it generates transparent images from text, edits existing transparent layers, and extracts a subject from an RGB photo as an RGBA layer, all in one model. FLUX.2 and Nano Banana Pro do not output an alpha channel; you need a separate background-removal pass.

Quick Answer

Qwen-Image-2.1 vs FLUX.2 dev vs GPT Image 2.5 vs Nano Banana

Published: September 21, 2026 • Updated: October 28, 2026

Last verified: October 27, 2026. (First published September 21, 2026.)

What changed as of October 2026

Qwen-Image-2.1-Turbo is out: an accelerated checkpoint of the same 7B model that generates and edits in 8 denoising steps instead of 40, loads with the same QwenImage21Pipeline, and keeps the RGBA and multi-reference features. Same Qwen research licence.
Black Forest Labs released FLUX 3 Image Klein on October 15, 2026: a 5.6B open-weights model (base + step-distilled, FP8/INT8/NVFP4 builds) for text-to-image, editing and up to 10 reference images, with native 4K and bounding-box placement control, under the FLUX Kommunity License. BFL says it leads its own blind evaluations of open models — a vendor claim, like Qwen’s. It is now the obvious small open-weight rival to Qwen-Image-2.1; FLUX.2 [dev] is no longer BFL’s newest open model.
Google added Nano Banana 2.1 (gemini-nano-banana-2.1), an update to Nano Banana 2, at about $0.034 per 1K, $0.050 per 2K and $0.113 per 4K image — now the cheapest route to 4K on Google.
Price fix: this page listed Nano Banana Pro at ”~$0.039 (1K)”. Google’s price for Nano Banana Pro is $0.134 per 1K or 2K image and $0.24 per 4K image; $0.039 was the older Nano Banana (Flash) rate. Corrected below and in the corrections log.
Independent results for Qwen-Image-2.1 are still thin: Alibaba has published only its own Qwen-Image-Bench score, and we could not confirm a third-party leaderboard placement at the primary source on October 27, 2026. The “vendor numbers” caveat stands.

Side by side

	Qwen-Image-2.1	FLUX.2 [dev]	GPT Image 2.5	Nano Banana Pro
Maker	Alibaba Qwen	Black Forest Labs	OpenAI	Google DeepMind
Released	September 20, 2026	November 25, 2025	September 8, 2026	Late 2025 (Gemini 3 Pro Image)
Weights	Open (research licence)	Open (non-commercial licence)	Closed	Closed
Size	7B DiT + Qwen3-VL 8B encoder	32B	Undisclosed	Undisclosed
Native resolution	2048x2048 (up to 2752x1536 wide)	Up to 4MP editing	Arbitrary up to 3840x2160	Up to 4096x4096
Transparent RGBA output	Yes, native	No	Yes, native	No
Reference images	Up to 10	Up to 10	Up to 16	Up to 14
Local edit control	Circles, painted marks, masks	Prompt + references	Multi-turn prompt edits	Conversational edits
Runs on consumer GPU	Yes (~24GB, RTX 3090 class)	fp8 build on RTX 40/50	No (API)	No (API)
API price	Free to self-host; Pro and Turbo APIs live on Alibaba Cloud Model Studio since October 9, 2026	fal / Replicate per-image; BFL API	Per token ($8 in / $30 out per 1M image tokens), ~$0.21 per 1024² at max quality	$0.134 (1K/2K) – $0.24 (4K) per image
Commercial use	Separate licence from Alibaba	BFL commercial licence	Yes	Yes
Independent benchmarks	None confirmed (Oct 27, 2026)	Yes	Yes	Yes

Verified September 21, 2026; re-verified October 27, 2026 against the Qwen Hugging Face model card and GitHub repo, BFL’s FLUX.2 announcement, and OpenAI’s and Google’s image pricing pages.

What Qwen-Image-2.1 changes

Alibaba framed the release around “compact and efficient”: a 7B single-stream DiT with mixed-granularity attention and prefix KV-cache reuse, so multi-reference edits do not re-encode every reference each step. Four things are new relative to the open-image field:

Native transparency. One model generates regular or RGBA images from text, edits existing transparent layers (change the text on a sticker without touching the alpha), and extracts a subject from an ordinary photo as a transparent layer. The recommended prompt literally states “This is an RGBA image with transparency… the background is transparent.”
Ten references. Group photos assembled from individual portraits, virtual try-on, and room design from furniture references, with identity preservation for people and products.
Annotation-driven local edits. Circle a region or paint a mark on the input and describe the change; no separate inpainting model.
Typography and portrait lighting improvements — Alibaba’s showcase leans on legible text rendering, historically the gap between open and closed models.

Setup is pip install transformers>=5.17 plus diffusers from git; QwenImage21Pipeline handles text-to-image, editing and RGBA in one class, and enable_model_cpu_offload() gets it onto smaller cards. Supported aspect presets run from 1:1 (2048²) to 16:9 (2752x1536).

The catch is the licence. Qwen-Image 1.0 (August 2025) was Apache-2.0; Qwen-Image-2.1 ships under the Qwen Research License Agreement — non-commercial only, separate grant for commercial use, “Built with Qwen” attribution. Alibaba’s closed Qwen-Image-3.0 (July 21, 2026; 4.5K-token prompts, 10px text, 12 languages) remains API/app-only with no weights, so 2.1 is the open line, and the open line just got a more restrictive licence.

How the other three compare

FLUX.2 [dev] is still the reference open-weight model for quality: 32B, text-to-image plus multi-reference editing in one checkpoint, up to 4MP, strong typography and world knowledge. Its weakness is footprint — 32B in bf16 is a data-centre model, and consumer use depends on the NVIDIA/ComfyUI fp8 build. BFL’s open-core model means [dev] is non-commercial by default with paid commercial terms, while [pro] and [flex] are API-only and the small [klein] variants (January 2026) cover the fast/cheap tier. See GPT Image 2 vs MAI-Image-2 vs FLUX.2.

GPT Image 2.5 (September 8, 2026) is the instruction-following leader: arbitrary resolutions to 3840x2160, native transparent backgrounds, up to 16 references, and two API variants — Flare for speed, Sunburst for precision (which to pick). It is per-token billed, roughly $0.21 for a 1024² image at max quality and about $0.40 at 4K max.

Nano Banana Pro (Gemini 3 Pro Image) costs $0.134 per 1K/2K image and $0.24 at 4K (Google’s paid tier, October 2026) and is the best conversational editor, with web-grounded factual text for infographics; Nano Banana 2 is the high-volume sibling, and Nano Banana 2.1 (October 2026) is now the cheapest 4K option at about $0.113 per 4K image. Full breakdown in ChatGPT Images 2.5 vs Nano Banana Pro vs Midjourney V8.2.

Where each one wins

Transparent assets (stickers, product cut-outs, UI icons): GPT Image 2.5 if you want an API; Qwen-Image-2.1 if you want it local and free to run for research or internal tooling.
Self-hosted, best absolute quality: FLUX.2 [dev], until independent evals of Qwen-Image-2.1 land.
Self-hosted on one consumer GPU: Qwen-Image-2.1 (7B, or the 8-step Turbo checkpoint), FLUX 3 Image Klein (5.6B) or Z-Image-Turbo (6B); FLUX.2 [dev] only in fp8.
Multi-person composites from references: Qwen-Image-2.1 and GPT Image 2.5 (10 and 16 references respectively) — FLUX.2 also handles 10 but without annotation-guided local edits.
Cheapest 4K at volume: Nano Banana 2.1 (~$0.113 per 4K image); Nano Banana Pro ($0.24) if you need its quality.
Commercial product, no legal friction: GPT Image 2.5, Nano Banana Pro, or FLUX.2 with a BFL commercial licence. Not Qwen-Image-2.1 without Alibaba’s grant.

The honest caveat

Alibaba’s claim that a 7B model beats “most closed models” rests on Qwen’s own benchmark, and the-decoder and others noted independent results were pending on release day. Multi-reference identity preservation and RGBA edge quality are exactly the properties vendor benchmarks flatter. If you are choosing for production, run your own 50-prompt bake-off; if you are choosing for experimentation, the 24GB footprint and the RGBA pipeline make Qwen-Image-2.1 the most interesting open image release since FLUX.2.

For the wider open field — Z-Image, HunyuanImage 3.0, FLUX.2 [klein] — see best open-weight image models 2026.