ModelsQwen3-VL 235B A22B Thinking: When Images Need Reasoning
The reasoning edition of Qwen's largest vision model leads on visual maths but costs about twice as much per token as Instruct. Which image tasks justify that.
Most of what people ask a vision model to do is perception: read this invoice, describe this screenshot, find the button. Reasoning does not help much there, and paying for a thinking trace on an OCR job is money spent on nothing. A smaller set of image tasks are genuinely multi-step: a geometry problem drawn on a whiteboard, a chart whose conclusion needs two calculations, a circuit diagram you want traced. That second set is what Qwen3-VL-235B-A22B-Thinking is for.
It was released alongside the Instruct edition on 23 September 2025, under Apache-2.0, with the same architecture and checkpoint size. The difference is post-training aimed at reasoning, and the model writes a reasoning trace before it answers.
Where thinking measurably helps
The Qwen3-VL technical report singles out the Thinking edition on multimodal maths: it claims state-of-the-art results on MathVista-mini, MathVision and MathVerse-mini, and the highest score in its comparison on MMStar, at 78.7. On 3D grounding, the report says the Thinking variant beats Gemini 2.5 Pro by 5.2 points on SUN RGB-D.
The text-only scores in the same report show how large the reasoning gap to Instruct is:
- AIME25: 89.7 for Thinking, 74.7 for Instruct
- LiveCodeBench v6: 70.1 against 54.3
- MMLU-Pro: 83.8 against 81.8
A fifteen-point jump on competition maths is a different kind of difference from two points on a knowledge test. Knowledge barely moves; multi-step problem solving moves a lot. Expect the same split on images.
Where it does not
The report is candid that the Instruct model marginally outperforms Thinking on OCR parsing benchmarks. That fits the general pattern of reasoning models: when the task is reading rather than working something out, a longer trace adds latency and cost and occasionally talks the model out of what it saw.
So the practical rule is to split by task, not by model preference. Extraction, captioning, screen reading and transcription go to the 235B Instruct edition. Charts that need arithmetic, diagrams that need tracing, visual puzzles and STEM questions with figures go here.
Against the text-only reasoning model
If your problems are text, the vision model is not the strongest Qwen option. Qwen's published figures for the text-only Qwen3-235B-A22B-Thinking-2507 are 92.3 on AIME25, 74.1 on LiveCodeBench v6 and 84.4 on MMLU-Pro, a few points above the VL Thinking model on each. Both sets sit in the same table of the Qwen3-VL report, but they are still vendor-run, so read the gap as indicative.
The trade-off is operational. One VL deployment that is slightly behind on text may beat running two models and a router, especially at low volume. At high volume with mostly text traffic, the dedicated text model is the better default and the vision model handles only the requests that carry images.
What a reasoning trace costs on images
OpenRouter listed this edition at $0.40 input and $4.00 output per million tokens at the time of writing, against $0.21 and $1.90 for Instruct. That is roughly double per token, before counting the trace. Since the trace is output, the real cost per answer can easily be several times Instruct's for the same image.
Qwen's evaluation settings show how long those traces get: 40,960 output tokens for Thinking in its repository settings, and up to 81,920 for the hardest maths and code benchmarks in the report. You will rarely need that much on a single chart, but you should set max_tokens deliberately per task class. The token cost of reasoning models covers how to measure this before it shows up in a bill.
Recommended sampling for Thinking is temperature 0.6, top-p 0.95, top-k 20 and no presence penalty, which differs from the Instruct settings. Copying Instruct's presence penalty of 1.5 into a Thinking deployment is an easy mistake.
Context and hosting notes
Qwen gives the native context as 256K tokens, extendable towards 1M with YaRN, where it recommends a scaling factor of 2 or 3 rather than 4. Hosted endpoints do not always serve the full native window: OpenRouter's listing showed 131,072 tokens at the time of writing. Check the limit your provider actually exposes before designing around long video or large document batches.
Self-hosting has the same requirements as the Instruct edition: roughly 470 GB of 16-bit weights by simple arithmetic, multi-GPU tensor parallelism, and an FP8 checkpoint for H100-class hardware.
A decision rule
- Label a sample of your real image requests as perception or reasoning.
- Run the reasoning ones through both 235B editions and score correctness, not fluency.
- If Thinking wins by a margin that matters, route that class to it and keep the rest on Instruct.
- If it does not, you have saved the trace cost on every request. Model routing and fallbacks covers the plumbing.
Common questions
Is Qwen3-VL-235B Thinking better than Instruct at OCR?
No. Qwen's technical report says the Instruct edition marginally outperforms Thinking on OCR parsing benchmarks. Use Thinking for multi-step visual reasoning such as maths, charts and diagrams.
What sampling settings does Qwen recommend for Qwen3-VL Thinking?
Temperature 0.6, top-p 0.95, top-k 20 and a presence penalty of 0. The Instruct edition uses different settings, including a presence penalty of 1.5, so do not copy one configuration into the other.
Should I use Qwen3-VL Thinking for text-only reasoning?
Only if one deployment matters more than a few points of accuracy. Qwen's text-only Qwen3-235B-A22B-Thinking-2507 scores somewhat higher on AIME25, LiveCodeBench and MMLU-Pro.
