ModelsQwen3-VL 235B A22B Instruct: The Open Vision Flagship
Qwen's largest vision-language model reads documents, screens and hours of video without the usual drop in text quality. What it does well and what it costs to run.
The standard complaint about vision-language models is that adding eyes costs the model some of its brain. A text model gets an image encoder bolted on, the combined model is fine-tuned on image data, and its plain-text reasoning quietly gets worse. Qwen3-VL-235B-A22B-Instruct is a useful counter-example, and the numbers that show it are worth looking at before the vision ones.
Qwen released the 235B-A22B pair, Instruct and Thinking, on 23 September 2025 under Apache-2.0. The Instruct edition answers directly without a reasoning trace, which makes it the default choice for most document, screenshot and video pipelines.
Text quality did not take the usual hit
The Qwen3-VL technical report gives text-only scores for the model and sets them beside Qwen's text-only Qwen3-235B-A22B-Instruct-2507 in the same table:
- MMLU-Pro: 81.8 for the VL model, 83.0 for the text model
- AIME25: 74.7 against 70.3
- LiveCodeBench v6: 54.3 against 51.8
These are vendor-run numbers, and small differences in either direction are noise. The point is that there is no large gap. Qwen's model card claims text understanding on par with pure LLMs, and the report's numbers support that. In practice it means one deployment can handle both the screenshot and the follow-up question about the code in it, rather than routing between a vision model and a stronger text model.
What the vision side is built for
The model card lists the targets plainly: operating desktop and mobile interfaces as a visual agent, generating HTML, CSS, JavaScript and Draw.io diagrams from images, 2D and 3D grounding, OCR in 32 languages with better handling of blur, tilt and low light, and long-video understanding with timestamp-level event localisation.
The technical report backs several of those with specifics. It says the Instruct model scored 89.3 and 88.9 on the two MMBench sets and 79.2 on RealWorldQA, the highest in its comparison; that it set a new state of the art in the report's document and OCR comparison; and that on video it performs on par with Gemini 2.5 Pro, matching or beating it on long-video tasks. In a video needle-in-a-haystack test it found the inserted frame with 100% accuracy in videos up to 30 minutes, about 256K tokens, and 99.5% when extended with YaRN to around 1M tokens, roughly two hours of video at one frame per second.
Three architectural changes sit behind that: interleaved MRoPE for positions across time, width and height; DeepStack, which feeds several levels of vision-encoder features into the language model rather than just the last one; and explicit text timestamps for video frames. Multimodal models explained covers the general mechanism of turning images into tokens.
Vision tokens are the cost you will underestimate
Qwen's repository example sets a budget of 256 to 1,280 visual tokens per image, scaled by resolution, and 256 to 16,384 per video. Those land in your input bill. A pipeline that sends ten full-resolution page scans per request is sending more than 10K tokens of images before any text.
OpenRouter listed the model at $0.21 input and $1.90 output per million tokens at the time of writing, with cached input cheaper. For document extraction that is usually input-dominated, so the resolution setting is the main cost lever. Run your documents at two resolutions and check whether accuracy actually drops before paying for the larger one.
The serving bill
Hugging Face lists the checkpoint at 236B parameters. At 16-bit that is roughly 470 GB of weights by simple arithmetic, and Qwen's serving examples use eight-way tensor parallelism for the FP8 checkpoint in vLLM and four-way in SGLang. Qwen also publishes an FP8 checkpoint, which halves the weight memory but requires NVIDIA H100-class hardware or newer.
About 22B parameters are active per token, so once it is loaded it is far cheaper per token than a dense model of the same size. Loading it is the hard part. For most teams that makes this a hosted model, with self-hosting reserved for cases where documents cannot leave your network. The Qwen3-VL 30B A3B Instruct is the realistic alternative when they cannot.
Recommended sampling for the Instruct edition is temperature 0.7, top-p 0.8, top-k 20 and a presence penalty of 1.5, with outputs up to 32,768 tokens.
Instruct or Thinking
The Qwen3-VL 235B Thinking edition scores much higher on reasoning: 89.7 on AIME25 against 74.7, and 70.1 on LiveCodeBench against 54.3. But the report notes that the Instruct model slightly outperforms it on OCR parsing, and it does not spend tokens on a trace. For extraction, captioning, screen reading and video summarisation, Instruct is the right default. Use Thinking for visual maths, diagram reasoning and charts that require several steps to interpret.
A test plan before you commit
- Collect thirty real inputs: your worst scans, your densest screenshots, one long video if you need video.
- Write the expected output for each by hand, including the JSON shape if you extract structured data.
- Run them through this model and the 30B sibling at the same resolution, and score field-level accuracy rather than overall impressions.
- Record visual tokens per input so you can price the result.
Vision-language models compared has a longer list of what to probe, including the failure modes that only appear on rotated or low-contrast pages.
Common questions
Is Qwen3-VL-235B-A22B-Instruct open weights?
Yes. Qwen publishes it on Hugging Face under Apache-2.0, along with an FP8 checkpoint. The permissive licence allows commercial use and fine-tuning without a separate agreement.
How long a video can Qwen3-VL-235B handle?
Its native context is 256K tokens, about 30 minutes of video at one frame per second in Qwen's test. With YaRN extension to around 1M tokens Qwen reports 99.5% needle-retrieval accuracy on roughly two hours of video.
Is Qwen3-VL worse at text than the text-only Qwen3 models?
Not by much. Qwen's technical report puts the 235B Instruct at 81.8 on MMLU-Pro, against 83.0 for the text-only Qwen3-235B-A22B-Instruct-2507, and slightly ahead on AIME25 and LiveCodeBench.
