Qwen3 VL 32B

qwenqwen/qwen3-vl-32b

Qwen3 VL 32B — mid-range vision-language model with tool use for multimodal workflows.

Hugging Face model card (Instruct) ↗Qwen3-VL GitHub repository ↗Qwen3-VL technical report (arXiv 2511.21631) ↗Qwen Chat demo ↗

Qwen3-VL-32B is Alibaba Qwen's flagship dense vision-language model in the Qwen3-VL family, released on October 21, 2025. It pairs a 32B-parameter language backbone with an enhanced ViT image encoder and three key architectural upgrades — Interleaved-MRoPE for temporal-spatial positional encoding across interleaved image/video/text sequences, DeepStack for multi-level ViT feature fusion, and Text–Timestamp Alignment for precise video temporal grounding. Available in two editions: Instruct (instruction-tuned, standard completion) and Thinking (reasoning-enhanced, with step-by-step CoT traces), both accessible on TheRouter through the single model ID qwen/qwen3-vl-32b.

For TheRouter users, Qwen3-VL-32B is the recommended mid-range vision-language model for workloads that need serious multimodal reasoning without the cost of the 235B MoE flagship. At $0.336 / $1.008 per million input/output tokens it handles image QA, OCR in 32 languages, GUI agent tasks, video understanding (native 256K context, expandable to 1M), chart and table extraction, and STEM visual reasoning. The Thinking edition's chain-of-thought output makes it practical for educational, scientific, and math-heavy pipelines that previously required a separate text reasoning model.

Best for

• Document intelligence and OCR — 32-language OCR with low-light, blur, and tilt robustness; structured extraction from invoices, contracts, tables, and long-document PDFs with native 256K context
• GUI and visual agent workflows — recognises UI elements, understands their functions, invokes tools, and completes tasks on PC/mobile screens; strong for screenshot-based agentic automation
• STEM / math visual reasoning — use the Thinking edition for competition-level math problems embedded in images, scientific diagrams, and multi-step causal analysis; outperforms Qwen2.5-VL-72B on MathVista and MathVision
• Long video comprehension — second-level indexing and full recall across hours-long video; temporal grounding via Text–Timestamp Alignment for retrieval, summarization, and QA over video archives
• Multimodal coding — generate Draw.io diagrams, HTML/CSS/JS, and structured data files from image or video inputs; useful for design-to-code and reverse-engineering UI screenshots

Reach for something else if

• Pure text tasks with no visual input — the 32B language backbone is capable, but qwen/qwen3-32b at lower price is the right choice for text-only workloads
• Highest-scale batch vision workloads where per-token cost is critical — consider qwen/qwen3-vl-8b for lighter tasks or qwen/qwen3-vl-235b (MoE, 22B active) for maximum quality at fleet scale
• Audio or native video generation — Qwen3-VL-32B is image+text-in / text-out only; it cannot produce audio, video, or rendered image outputs

Context Length

33K

Max Output

Input Priceper 1M tokens

$0.1728/ 1M tokens

Output Priceper 1M tokens

$0.6912/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

Type	Rate
Input	$0.1728 / 1M tokens
Output	$0.6912 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choicestop

Specifications

Release date	2025-10-21github.com ↗	verified
Architecture	Dense transformer; ~32B parameters; enhanced ViT image encoder with Interleaved-MRoPE and DeepStackarxiv.org ↗	verified
Input modalities	Text, images, video (interleaved sequences supported)huggingface.co ↗	verified
Native context length	256K tokens (expandable to 1M)github.com ↗	verified
OCR languages	32 languageshuggingface.co ↗	verified
License	Apache 2.0huggingface.co ↗	verified
Training cutoff	Not publicly disclosed	unknown

Benchmarks

Benchmark	Score	Source
MMMU	76.0%	www.emergentmind.com ↗
MathVista_mini	83.8%	www.emergentmind.com ↗
MathVision	63.4%	www.emergentmind.com ↗
MMLongBench-Doc	54.6%	www.emergentmind.com ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "qwen/qwen3-vl-32b",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Vision chat completion

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-vl-32b",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "image_url", "image_url": {"url": "https://therouter.ai/assets/vision-sample.png"}},
        {"type": "text", "text": "Describe what you see in this image."}
      ]
    }]
  }'

Similar models

Cross-provider sibling models

qwen/qwen3-vl-235b

MoE flagship (22B active) in the same Qwen3-VL family; higher quality ceiling for the most demanding visual reasoning tasks at fleet scale

qwen/qwen3-vl-8b

Compact sibling; choose when per-token cost matters more than peak reasoning quality

google/gemini-2.5-flash

Closed-model alternative with strong vision and native thinking; benchmark against Qwen3-VL-32B for latency and quality trade-offs on your target tasks

News & changes

2025-11-27

Qwen3-VL technical report published on arXiv (2511.21631)

The Qwen Team released the full 42-page technical report covering the Interleaved-MRoPE positional encoding scheme, DeepStack vision-language alignment, text-timestamp video grounding, and comprehensive benchmark results across the Qwen3-VL dense (2B/4B/8B/32B) and MoE (30B-A3B/235B-A22B) family.

re-authored by TheRouterarxiv.org ↗

2025-10-21

Qwen3-VL-32B-Instruct and Thinking editions released

Alibaba Qwen released the Qwen3-VL-32B in both Instruct and Thinking editions on Hugging Face, completing the mid-tier rollout of the Qwen3-VL family. The Thinking edition adds chain-of-thought reasoning to the vision pipeline, enabling STEM and math-heavy visual tasks that previously required a separate text reasoning model.

re-authored by TheRoutergithub.com/QwenLM ↗

Frequently asked

What is the difference between Qwen3-VL-32B Instruct and Thinking on TheRouter?

Instruct is optimised for direct, concise responses to visual questions — lower latency, no reasoning trace overhead. Thinking produces step-by-step chain-of-thought reasoning before its final answer, increasing token count and cost but significantly improving accuracy on STEM, math-embedded-image, and multi-step visual reasoning tasks. On TheRouter the qwen/qwen3-vl-32b model ID routes to the Instruct edition by default; check the TheRouter docs for the Thinking edition routing parameter.

huggingface.co ↗

Can Qwen3-VL-32B process video, or only images?

Qwen3-VL-32B natively supports video input in addition to images and text, with a 256K-token context window expandable to 1M. It provides second-level temporal indexing and full recall across long videos via the Text–Timestamp Alignment architecture. Video support availability through TheRouter's /v1/chat/completions endpoint may differ from direct API access — check the current TheRouter model card for supported input types.

github.com ↗

How does Qwen3-VL-32B compare to Qwen2.5-VL-72B?

Qwen3-VL-32B-Instruct matches or exceeds Qwen2.5-VL-72B on MMMU (76.0 vs 77.7) and surpasses it on MathVista (83.8 vs 79.4) and MathVision (63.4 vs 64.3) with fewer than half the parameters. The Thinking edition widens the gap further on math-heavy benchmarks (MMMU 78.1, MathVista 85.9, MathVision 70.2). For most production multimodal workloads, Qwen3-VL-32B offers a better cost-quality trade-off than the prior-generation 72B model.

re-authored by TheRouterwww.emergentmind.com ↗

Can I use Qwen3-VL-32B from Cursor, Cline, or another OpenAI-compatible client?

Yes. Set your client's base URL to https://api.therouter.ai/v1, configure your TheRouter API key, and set the model to qwen/qwen3-vl-32b. Image inputs follow the standard OpenAI vision format (image_url content parts). Streaming and tool calling are supported. Clients that only send text will work normally — the vision capability is additive.