Alibaba claims new image model beats Google’s Nano Banana 2.0: Qwen-Image-2.1, a 7B open-weight model, scores higher on Alibaba’s own benchmark but remains unverified by independent tests.

Alibaba claims new image model beats Google’s Nano Banana 2.0

Alibaba claims new image model beats Google’s Nano Banana 2.0, after its Qwen team released Qwen-Image-2.1 on September 20, 2026. The new 7‑billion‑parameter open‑weight model unifies text‑to‑image generation and image editing in a single architecture and, on Alibaba’s own Qwen-Image-Bench, scores slightly higher than Google’s closed Nano Banana 2.0 (Gemini 3.1 Flash Image).

The claim has drawn attention because Qwen-Image-2.1 is one of the few open-weight image models that reportedly outperforms major proprietary systems on at least one published benchmark. However, the comparison relies entirely on Alibaba’s internal evaluation; independent leaderboards have not yet verified the results.

What Qwen-Image-2.1 is

Qwen-Image-2.1 is a unified image generation and editing model with:

• A 7B-parameter visual generator, down from 20B in the original Qwen-Image, designed to be more compact and cost‑effective.
• Native support for transparent (RGBA) image output, so alpha channels are built into the base model rather than added as a separate feature.
• Ability to take up to 10 reference images and compose them into a single coherent output, supporting complex multi‑image editing workflows.
• Integrated text-to-image and image-editing capabilities in one model, rather than separate generation and editing systems.

Alibaba describes it as “the most balanced and cost-effective image generation model in the Qwen-Image series,” emphasising speed for multi-image inputs and a lighter footprint for deployment.

Benchmark claims vs Nano Banana 2.0

On Alibaba’s Qwen-Image-Bench, Qwen-Image-2.1 achieves an overall score of 60.28, compared with:

• Nano Banana 2.0: 59.82 (0.46 points lower).
• GPT Image 1.5: 59.65.
• Higher-scoring closed models such as GPT Image 2.5 Sunburst at 67.01, which still lead the chart.

This places Qwen-Image-2.1:

• Above all open-weight image models on the chart, including larger models such as FLUX 2 Max (32B).
• Slightly ahead of Google’s Nano Banana 2.0 on this specific benchmark.
• Behind several top closed models, especially the latest OpenAI variants.

Alibaba has not published a detailed score table on Hugging Face or in the official blog; the numbers appear in presentation slides and summary posts. Independent evaluators have not yet reproduced the results on public arenas.

Availability and license

Qwen-Image-2.1 weights are available on:

• Hugging Face (Qwen/Qwen-Image-2.1).
• ModelScope (Alibaba’s model platform).
• GitHub repositories with example code and integration notes.

However, the model is released under the Qwen Research License, which:

• Allows free use for non-commercial research.
• Requires a separate commercial license for product or business use.
• Replaces the earlier Apache 2.0 license used for some prior Qwen models, sparking discussion about openness and commercial terms.

Tools such as Diffusers and ComfyUI added support for Qwen-Image-2.1 around launch, enabling local and cloud-based workflows for users who obtain the weights.

How it compares in broader rankings

Outside Alibaba’s own benchmark, Qwen-Image-2.1’s position is less clear:

• On OpenArt’s Arena leaderboards for image tasks, other models such as Seedream 5.0 Pro, GPT Image 2, and Nano Banana Pro lead specific categories like film-oriented imagery, graphic design, and e‑commerce images.
• Qwen-Image-2.1 does not yet appear prominently on these third-party boards, and no independent score directly pits it against Nano Banana 2.0 outside Alibaba’s evaluation.

This means the “beats Nano Banana 2.0” claim should currently be treated as Alibaba’s internal assertion, not an independently confirmed superiority across all image tasks.

Summary: Alibaba claims new image model beats Google’s Nano Banana 2.0 with Qwen-Image-2.1, a 7B open-weight model that scores 60.28 vs 59.82 on Alibaba’s Qwen-Image-Bench. The model supports native transparency, 10-reference editing, and unified generation/editing, but the benchmark advantage is based on Alibaba’s own evaluation and has not yet been independently verified.