Alibaba's Qwen team released Qwen-Image-2.1 today with open weights on Hugging Face and ModelScope: one model that generates images from text, edits them from up to ten reference pictures, and outputs native transparency, at 2048×2048 out of the box. It is the first open-weight release in the line since December. It is also the first one that is not Apache 2.0, and that second fact will matter to more of our readers than the first.
What it is
The original Qwen-Image from August 2025 was a 20-billion-parameter dual-stream diffusion transformer that made its name on text rendering, especially Chinese. Qwen-Image-2.0 followed in February 2026 as an API-only product; seven months later its weights still have not been published. Version 2.1 goes back to open weights, and it is a smaller, different design:
| Qwen-Image (2025) | Qwen-Image-2.1 | |
|---|---|---|
| Diffusion transformer | 20B, dual-stream MMDiT | 7B, 32 single-stream layers |
| Text and image conditioning | Qwen2.5-VL | Qwen3-VL 8B |
| VAE | 16-channel RGB | 64-channel RGBA, 16× compression |
| Native resolution | 1328×1328 class | 2048×2048, seven aspect ratios to 2752×1536 |
| Editing | separate Qwen-Image-Edit model | same model, up to 10 reference images |
| Transparency | separate Layered model | native |
| Licence | Apache 2.0 | Qwen Research License |
The architectural idea is a single stream with block-causal attention: text tokens attend causally, image tokens attend bidirectionally within their block, and the prompt plus any reference images are encoded once at the first denoising step and reused as a cached prefix for the other 39. That is the same trick as a language model's KV cache, applied to a diffusion model, and it is why an edit with several reference images does not cost several times a plain generation.
The post also folds in two capabilities that used to be separate products: Qwen-Image-Layered, the transparency model from December 2025, is now a prompt away inside the main model, and editing has grown task coverage the earlier line did not advertise, including panoramas from a single selfie, infographics from a product photo and storyboards from a three-view character sheet. All of those are shown as examples, none as measurements.
Two companion models ship with it: prompt rewriters for text-to-image and for editing, both fine-tuned from Qwen3.5-VL 9B, which expand a short prompt into the long descriptive kind the model was trained on and pick an aspect ratio for you. Day-zero support is unusually broad. Diffusers has a pipeline class, ComfyUI has native nodes and two template workflows, vLLM-Omni serves it with FP8 and CUDA graphs, SGLang has a cookbook down to consumer cards, LightX2V has tuned configs for the RTX 5090, and FlagOS publishes prebuilt images for eight non-Nvidia chip platforms.
The licence, read in full
The earlier line, Qwen-Image, Qwen-Image-2512 and Qwen-Image-Edit-2511, is Apache 2.0. Qwen-Image-2.1 is released under the Qwen Research License Agreement, dated today. We read it. The operative clauses:
- Rights are granted "FOR NON-COMMERCIAL PURPOSES ONLY", and the agreement defines non-commercial as "for research or evaluation purposes only".
- "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us", with an email address for requests.
- Redistribution and derivatives are allowed, with an attribution notice naming Hangzhou Tongyi Laboratory.
So a designer generating product shots for a client, a shop making banners, an agency building a tool on it: all of that needs a commercial licence that does not yet have a published price. This is the pattern Qwen used for its largest language models and it is new for the image line. If your workflow is built on the Apache-licensed Qwen-Image-2512 or Edit-2511, nothing about those changes; they remain usable for anything. If you were planning to swap them for 2.1, the licence is the first thing to send to whoever signs contracts.
What it needs to run
The README does not say. The vendor's own inference recipes do, and the numbers are worth a table because the headline "7B" undersells them: the text encoder is bigger than the diffusion model.
| component | BF16 weights |
|---|---|
| Qwen3-VL 8B text encoder | 17.5 GB |
| DiT, 7B | 14.2 GB |
| RGBA VAE | 1.4 GB |
| total | ~33 GB |
| configuration, 1024×1024 | peak VRAM | source |
|---|---|---|
| all BF16, resident, 40 steps | 34.0 GB | vLLM-Omni on GB300 |
| all BF16, 50 steps | 40.0 GB | vLLM-Omni on GB200 |
| text encoder FP8, DiT BF16 | 27.5–28.1 GB | vLLM-Omni on GB300 |
| everything FP8 | 33.4 GB | vLLM-Omni on GB200 |
Fully resident in BF16, it does not fit a 32 GB RTX 5090, and SGLang's cookbook says so directly: the 5090 and 4090 "exceed single-GPU capacity and require offload". There are two ways out, and both shipped on day one.
Offload the text encoder. It runs once per prompt, so it can live in system RAM and stream through the card layer by layer while the 14.2 GB DiT and the VAE stay resident. That is SGLang's recommended layout for both consumer cards. LightX2V goes further with an FP8 DiT and publishes the only consumer-card timing we have found:
| RTX 5090, FP8 DiT, 40 steps, 1024×1024 | seconds |
|---|---|
| text to image | 5.93 |
| image edit | 7.14 |
A datacentre GB300 in BF16 does the same image in 3.3 to 4.5 seconds, so the consumer card is not far behind once memory is solved.
Or quantise everything. Comfy-Org's official repackage, published the same day, carries INT8 versions of both big components alongside the BF16 originals, and the file sizes settle the question for the ComfyUI crowd:
| component | BF16 | INT8 (convrot) | smaller still |
|---|---|---|---|
| diffusion transformer, 7B | 13.3 GiB | 6.8 GiB | NVFP4 3.9 GiB |
| text encoder, Qwen3-VL 8B | 16.3 GiB | 8.7 GiB | W4A8 5.9 GiB |
| VAE | 0.6 GiB | 0.6 GiB | 0.6 GiB |
| total | 30.2 GiB | 16.1 GiB | 10.4 GiB |
At 16.1 GiB the entire pipeline sits on a 5090 with 15 GiB to spare for activations at 2K, and it fits a 24 GB 4090 resident with nothing streamed. The 10.4 GiB combination reaches 16 GB cards. The INT8 prompt rewriters are in the same repository at 8.8 GiB each, though they run before generation and never need to share the card with it. What no one has published is a quality comparison between the INT8 and BF16 outputs, so treat the small files as a fit, not yet as a free lunch.
Every published speed and peak-memory figure is at 1024×1024. Native output is 2K, four times the pixel area, and nobody has published a peak-memory number at that size. That is the measurement we intend to take. A 128 GB GB10 box runs the whole pipeline resident in BF16 with room to spare, which makes this the same capacity-versus-bandwidth trade as every other model we have run this year, only with a 5090 that now costs three times its list price.
The benchmark is Qwen's own
The launch post carries exactly one comparison: a chart of Qwen-Image-2.1 against "other open-source and closed-source models" on Qwen-Image-Bench. That is the evaluation suite Qwen itself published in May 2026, so the vendor is grading its model on its own exam, and the scores appear as a figure rather than a table. The GitHub README and the Hugging Face model card contain no benchmark at all, and no independent evaluation exists six hours after release. No verified comparison against Flux, Seedream, GPT-Image or the 20B predecessor exists yet. The "improved typography, portrait lighting and fine details" in the announcement are the vendor's words, illustrated with the vendor's examples.
What we can say is what the design promises. A 7B transformer with a cached prefix is built for cheap editing rounds, not for winning a quality leaderboard against 20B models; native RGBA and ten-image conditioning are built for product and design workflows; and a research licence is built to be sold. Whether the quality holds up at 7B is exactly what a local test will show, and the text-rendering check will be in Serbian, both scripts, as it was for DeepSeek-V4-Flash.
Bottom line
- Open weights again, after 2.0 stayed behind an API: 7B DiT, Qwen3-VL 8B encoder, RGBA VAE, native 2K, editing and generation in one model.
- Not Apache any more. Research and evaluation only; commercial use needs a separate licence with no published price. The Apache-licensed 2512 and Edit-2511 models are unaffected.
- 33 GB of BF16 weights, 16 GiB in INT8. A 5090 runs BF16 with the text encoder offloaded, at about 6 seconds per 1K image with an FP8 DiT; Comfy-Org's INT8 repackage puts the whole pipeline on a 5090 or a 4090 with nothing streamed. Resident BF16 starts at 48 GB.
- No benchmarks we could verify. Vendor claims only, until someone measures it.
Related reading
- A Local Suno Alternative: MiniMax Music 3 on One RTX 5090 — the last generative model we put on the card, and the template for the test to come
- AI Workstation Comparison: RTX 5090 vs GB10 (HP ZGX) — why 33 GB of weights is a different problem on each
- Quantization Methods Compared: GGUF, AWQ, GPTQ, EXL2, NVFP4 — the formats that get this under 24 GB
- RTX 5090 Prices: 3× Launch in the US, 2.4× in Europe, and Still Climbing — the card this article assumes you have