Qwen3-VL
Alibaba's Apache 2.0 vision-language family from 2B to 235B with 256K context. Superseded by Qwen3.5 but still widely used.
In short. Qwen3-VL is Alibaba's 2025 family of models that read images and video as well as text, in sizes from 2B for small devices to 235B for data centers. Newer Qwen generations can already see, so this one is marked legacy, but its 4B and 8B versions are still downloaded in the millions and run on ordinary hardware.
Summary
Alibaba's Qwen team released Qwen3-VL in stages under the Apache 2.0 license: the 235B-A22B mixture-of-experts models on 23 September 2025, the 4B and 8B dense models on 15 October 2025, and the 2B and 32B models on 21 October 2025, with a 30B-A3B mixture-of-experts model in between. Every size has an Instruct edition and a reasoning-oriented Thinking edition, and Qwen also publishes FP8 and GGUF builds. The models take text, images and video, with a native 256K-token context that the card says can be expanded to 1M. The technical report is arXiv 2511.21631.
According to the model card, the family adds a visual agent that operates PC and phone screens, visual coding that turns images or video into Draw.io, HTML, CSS and JavaScript, stronger spatial perception with 2D and 3D grounding, long-video understanding with timestamp-grounded events, and OCR in 32 languages, up from 19. The architecture adds interleaved multi-axis rotary position embeddings, DeepStack fusion of multi-level vision features, and text-based timestamp alignment for video. The benchmark charts on the cards are Qwen's own.
We mark Qwen3-VL as legacy because Qwen's later generations, including Qwen3.5, take images and video natively in the main model, so new projects should start there. Qwen3-VL remains heavily used: the 8B Instruct model had about 7.0 million downloads and the 4B about 3.6 million in the 30 days to 11 November 2026, and Qwen's own Qwen3-VL-Embedding models are built on it. Its small sizes, mature tool support and official GGUF and FP8 builds are the reasons people still choose it.
Good for
- Reading images, charts, screenshots and video with a small model that runs on an ordinary GPU or laptop (2B to 8B)
- GUI agents that operate screens, and generating HTML or CSS from screenshots
- OCR and document parsing in 32 languages
- Long video and long document understanding with a native 256K context
Watch out for
- Marked legacy because Qwen3.5 and later models see images and video natively, so start new projects there
- Instruct and Thinking are separate downloads, and Thinking is the reasoning edition
- The benchmark charts are Qwen-reported
- The 235B-A22B model needs about 471 GB in BF16, so use the FP8 or GGUF builds or a multi-GPU server
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 2B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 2B BF16 | 2.13B | 4.9 GB | 2.45 GB | 1.47 GB | Laptop or phone |
| 4B BF16 | 4.44B | 10 GB | 5.11 GB | 3.06 GB | Laptop or phone |
| 8B BF16 | 8.77B | 20 GB | 10 GB | 6.05 GB | Laptop or phone |
| 30B-A3B BF16 | 31.1B 3B active | 71 GB | 36 GB | 21 GB | Gaming GPU |
| 32B BF16 | 33.4B | 77 GB | 38 GB | 23 GB | Gaming GPU |
| 235B-A22B BF16 | 236B 22B active | 542 GB | 271 GB | 163 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is Qwen3-VL?
- Qwen3-VL is Alibaba's 2025 family of models that read images and video as well as text, in sizes from 2B for small devices to 235B for data centers. Newer Qwen generations can already see, so this one is marked legacy, but its 4B and 8B versions are still downloaded in the millions and run on ordinary hardware.
- Can I use Qwen3-VL commercially?
- Yes. Qwen3-VL is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Qwen3-VL open source?
- Qwen3-VL is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Qwen3-VL on my own computer?
- Yes. The smallest variant, 2B, needs about 1.47 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Qwen3-VL released?
- Qwen3-VL was released by Alibaba Qwen on September 23, 2025.
- How long a context window does Qwen3-VL have?
- Qwen3-VL supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Qwen3-VL repository, release news (GitHub) github.com
- Qwen blog: Qwen3-VL qwen.ai
- Qwen3-VL-8B-Instruct model card (Hugging Face) huggingface.co
- Qwen3-VL technical report (arXiv 2511.21631) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Qwen3-VL-Embedding Alibaba Qwen | 2.13B – 8.14B | 32K | Apache 2.0 (Permissive) | 4.9 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |