Qwen-Image-2.1: Is There an API, and What Can You Use Today?

Is there a Qwen-Image-2.1 API?

No. As of September 21, 2026 there is no official Qwen-Image-2.1 API, and no sign of one:

Item	Qwen-Image-2.1 status
Open weights	Released September 20, 2026
Official hosted API	Not published
Public API model ID	Not published
Official price	Not published
Licence	Qwen Research License — non-commercial research and evaluation only
What exists today	Download weights, Diffusers pipeline, ComfyUI nodes, Hugging Face demo Space

Anyone quoting a "Qwen-Image-2.1 API price" is pricing a self-hosted GPU, not a vendor endpoint. That figure depends entirely on your own hardware and utilisation, and it is not comparable to a per-image API rate.

What is actually confirmed about Qwen-Image-2.1

Everything below comes from Qwen's own release materials and model card.

• Generation and editing share one set of weights. A single QwenImage21Pipeline does text-to-image and instruction-based editing, so you do not swap models to switch tasks. Text and reference images are encoded first, then denoised.
• A ~7B image backbone plus a vision-language encoder. The visual part is a 32-layer Single-Stream DiT at roughly 7B parameters, with text and reference images encoded by Qwen3-VL-8B. Counting the text encoder, the stack is on the order of 16B, and the official BF16 weights are about 33GB.
• Native transparency (RGBA). The VAE is a 64-channel, 16× spatial-compression RGBA autoencoder. It can generate images with an alpha channel directly, edit existing transparent layers, and cut subjects out of ordinary photos into transparent assets. Qwen's own prompting advice is to state it explicitly, for example: "This is an RGBA image with transparency. … The image has alpha channel and the background is transparent."
• Up to 10 reference images. One request can carry up to 10 conditioning images for group shots, virtual try-on or interior staging. Local edits accept circles, brush marks or a separate mask, with an emphasis on preserving identity for people and products. The ComfyUI nodes reserve slots up to image_16.
• Native 2K output. The common size is 2048×2048, and a 16:9 frame reaches roughly 2752×1536. The default step count is 40. Qwen highlights improved text layout, portrait lighting and fine detail.
• Efficiency work. Mixed-granularity attention and prefix KV cache reuse mean the prompt and reference images are encoded once and reused across denoising steps, which saves the most compute when editing with several reference images.
• Day-one ecosystem. Hugging Face Diffusers (QwenImage21Pipeline), ComfyUI (with separately packaged quantized weights, a minimum combination around 14GB), plus vLLM-Omni, SGLang and LightX2V.
• Prompt-rewrite models. Qwen also ships PE-T2I / PE-I2I prompt-rewriting models — roughly 9B Qwen3.5-VL fine-tunes.

What is not confirmed

• No hosted endpoint from anyone official. Qwen published weights, not a service.
• No official price — not per image, not per token.
• The benchmark is the vendor's own. Qwen reports about 60.28 on its in-house Qwen-Image-Bench, which its chart places first among open models, marginally ahead of Nano Banana 2.0 (59.82) and GPT Image 1.5 (59.65). That is a self-built benchmark, and independent replication was scarce on release day. Treat the ranking as a claim to test, not a result to quote.

The licence is the real decision point

This is the part most "Qwen-Image-2.1 is here" coverage skips. The weights are under the Qwen Research License, effective September 20, 2026: research and evaluation only, non-commercial. Running it inside a commercial product, a client deliverable, or a paid API of your own generally requires a separate commercial grant from Qwen.

That is a genuine change in posture. Earlier Qwen-Image releases were Apache 2.0, which is why so many tutorials assume "download and ship it". With 2.1, the download is not a commercial licence. Before you build anything customer-facing on these weights, read the licence text in the Hugging Face repo and, if you need commercial rights, ask Qwen directly.

Qwen-Image-2.1 vs the image APIs you can use commercially

	Qwen-Image-2.1	GPT Image 2 (gpt-image-2)	Nano Banana 2 (gemini-3.1-flash-image)	Nano Banana (gemini-2.5-flash-image)	Nano Banana Pro (gemini-3-pro-image)
Access	Open weights, self-host	API	API	API	API
APIMaster routes	Not hosted	5 live routes	3 live routes	Live	Live
APIMaster price, per 1K image	No API price (your GPU cost)	from $0.0075	from $0.0469	$0.03744	$0.12864
Reference price per image	—	$0.25	$0.067	$0.039	$0.134
Commercial use	No (research/evaluation licence)	Yes	Yes	Yes	Yes

Prices are APIMaster's live route prices, checked September 21, 2026, and are separate from the vendors' own reference rates; account adjustments may apply. Qwen-Image-2.1 has no API price because it has no API — only a download. Note also that native RGBA transparency is Qwen-Image-2.1's standout feature; whether a hosted route returns a true alpha channel varies by channel, so verify on the route you pick before relying on it.

What to use today

• Cheapest commercial image generation and editing: GPT Image 2 (gpt-image-2). At from $0.0075 per 1K image against a $0.25 reference, it is the lowest-cost route on the platform and covers both generation and reference-image editing. Start on the GPT Image model page.
• The closest alternative: Nano Banana 2 (gemini-3.1-flash-image). Google's current Flash image tier, from $0.0469 per 1K image against a $0.067 reference. See the Nano Banana model page.
• Nano Banana and Nano Banana Pro cover the lighter and premium ends of the same family at $0.03744 and $0.12864 per 1K image.

Qwen is not absent from the platform — its text models are listed and live: qwen3.8-max, qwen3.8-flash, qwen3.7-max and qwen3.7-plus, all reachable through one key on the Qwen API page. The image models are a different matter: because Qwen-Image-2.1 is non-commercial and has no hosted API, there is nothing we can route today. If Qwen opens a commercial licence or a hosted image endpoint, we will evaluate listing it — the same key you use now would then cover it with no new integration.

How to call a commercial image API

Create a key in the APIMaster console, set APIMASTER_API_KEY, and use the channel schema shown below. On these routes size is an aspect ratio and resolution selects the pricing tier. These conventions belong to the channel, so do not copy them into a direct vendor request.

curl "https://apimaster.ai/v1/images/generations" \
  -H "Authorization: Bearer $APIMASTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A studio product photograph of a red ceramic teapot on a teal table.",
    "size": "1:1",
    "resolution": "1K",
    "n": 1
  }'

Switching to the Google tier is a one-line change — set "model": "gemini-3.1-flash-image" and keep the same key and endpoint.

Preparing for a Qwen-Image-2.1 API

1. Keep the model ID configurable. If a hosted qwen-image endpoint ever appears, treat it as a config value rather than hardcoding a guessed string.
2. Read the licence first. Non-commercial today means no customer-facing use until Qwen grants commercial rights.
3. Benchmark on your own tasks. Record correctness, latency and cost per usable image on the route you actually use, so any future comparison is fair.
4. Verify what a route is serving. A model name alone cannot prove which model answered; check a channel before scaling it.

FAQ

Is Qwen-Image-2.1 released?

The weights are. Qwen open-sourced Qwen-Image-2.1 on September 20, 2026 for Hugging Face, GitHub, ModelScope and ComfyUI. There is no hosted API release.

Is there an official Qwen-Image-2.1 API or price?

No. Qwen has published no API model ID and no price for Qwen-Image-2.1. Any cost you see quoted is a self-hosting estimate, not a vendor rate.

Can I use Qwen-Image-2.1 commercially?

Not under the released licence. The weights use the Qwen Research License, limited to non-commercial research and evaluation; commercial use requires a separate grant from Qwen.

Does APIMaster host Qwen-Image-2.1?

No. APIMaster lists Qwen's text models (qwen3.8-max, qwen3.8-flash, qwen3.7-max, qwen3.7-plus) but not Qwen's image models, because there is no hosted API to route. That would change only if Qwen opens a commercial licence or a hosted endpoint.

So what should I use for commercial image generation?

GPT Image 2 at from $0.0075 per 1K image (about 97% off OpenAI's $0.25 reference), or Nano Banana 2 at from $0.0469 per 1K image. Both are live, both support editing, and both are billed per image on your APIMaster key.

Is the 60.28 benchmark score trustworthy?

It is Qwen's own in-house benchmark, presented in Qwen's launch chart as the top open model. Independent replication was limited on release day, so treat the number as a self-reported claim rather than a verified result.

What is Qwen-Image-2.1's standout feature?

Native RGBA transparency: a 64-channel RGBA VAE lets it generate and edit images with a real alpha channel, which is unusual among image models and hard to get from many hosted routes.

Sources and verification date

• Qwen-Image-2.1 weights and model card: huggingface.co/Qwen/Qwen-Image-2.1 — architecture, RGBA VAE, reference-image and resolution details.
• Qwen release code and nodes: github.com/QwenLM/Qwen-Image-2.1.
• Qwen release blog: qwen.ai/blog?id=qwen-image-2.1 — licence and benchmark framing.
• ModelScope mirror: modelscope.cn/models/Qwen/Qwen-Image-2.1.
• APIMaster live pricing catalogue and marketplace cards, checked September 21, 2026. Route availability and prices can change after publication.

Start with an image API that exists today

Qwen-Image-2.1 is worth watching — native transparency and 10-image editing are genuinely interesting. But it is not an API you can call, and its licence rules out commercial use. GPT Image 2 and Nano Banana 2 are live right now.

Create an APIMaster account, generate a key in the console, point your client at https://apimaster.ai/v1 and set the model to gpt-image-2 — from $0.0075 per 1K image, about 97% off the $0.25 reference. One OpenAI-compatible key also covers Nano Banana, Claude, GPT, DeepSeek and Qwen's text models, and every channel is fingerprint-checked before it is listed.