Qwen3 VL 235B A22B Thinking

qwen/qwen3-vl-235b-a22b-thinking
VisionToolsJSONReasoning
by Qwen · 2025-09-23

Qwen3-VL 235B-A22B Thinking — open-weight vision-language reasoning model, 235B total / 22B active params, 128k context.

ctx131.1K tokens
Max output41K
Inputtext + image + video
Outputtext
p50 TTFT4.14 s
INPUT$0.40/ 1M tokens
OUTPUT$4.00/ 1M tokens
p50 TTFT4.14 s7d
p95 TTFT10.00 s7d
TRAFFIC718.4Ktokens / 7d

Qwen3 VL 235B A22B Thinking is a large-scale multimodal language model from the Qwen family. It employs a mixture-of-experts architecture, where only 22 billion of its 235 billion parameters are…

What is Qwen3 VL 235B A22B Thinking?

Who is this model for?

What input modalities does it support?

How does the thinking mode work?

Code samples

Call from any SDK

OpenAI-compatible — keep the SDK you already use

  • OpenAI SDKhttps://api.orcarouter.ai/v1
from openai import OpenAI

client = OpenAI(
    base_url="https://api.orcarouter.ai/v1",
    api_key="$ORCAROUTER_API_KEY",
)

response = client.chat.completions.create(
    model="qwen/qwen3-vl-235b-a22b-thinking",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

Supported parameters

  • enable_search
  • enable_thinking
  • include_reasoning
  • logprobs
  • max_tokens
  • n
  • parallel_tool_calls
  • presence_penalty
  • reasoning
  • repetition_penalty
  • response_format
  • seed
  • stop
  • stream
  • stream_options
  • temperature
  • thinking_budget
  • tool_choice
  • tools
  • top_k
  • top_logprobs
  • top_p

Pricing

Input / 1M tokens$0.400
Output / 1M tokens$4.00
CurrencyUSD

Cost calculator

Tokens / month10MM
Input share70%%
Estimated / month $14.80

Estimate based on list price

Token & cost estimator

Input tokens: 20Cost per request: $0.002008

Estimate only — actual token counts depend on the provider's tokenizer.

Performance

p50 TTFT
4.14 s
Output speed
38.2 tok/s
p95 TTFT
10.00 s
Error rate
0%

Public benchmarks

Source: Design Arena

How it compares

Qwen3 VL 235B A22B Thinkingqwen/qwen3-max-previewQwen3.5 397B A17Bqwen/qwen3.5-plus
Input $/M$0.40$0.86$0.17$0.12
Output $/M$4.00$3.44$1.03$0.69
Context131K262K33K1.0M
Quality6/108/108/108/10
Compare side-by-sideCompare side-by-sideCompare side-by-sideCompare side-by-side

FAQ

What is the cost per token for Qwen3 VL 235B A22B Thinking?
Input tokens cost $0.40 per 1 million tokens; output tokens cost $4.00 per 1 million tokens. These are provider rates with zero markup from OrcaRouter.
What is the context window size?
The model has a context window of 131,072 tokens, which includes both input and output tokens. The maximum output is 40,960 tokens.
What are the main strengths of this model?
Its strengths include a mixture-of-experts architecture for efficient scaling, a built-in thinking mode for chain-of-thought reasoning, support for text, image, and video inputs, and a large context window. It is suitable for complex multimodal tasks.
How does it compare to OpenAI's GPT-4o?
Qwen3 VL uses MoE with 22B active parameters, while GPT-4o is dense. It is cheaper per token ($0.40/$4 vs $5/$15) and has an optional thinking mode. However, GPT-4o may have lower latency and different performance characteristics on specific benchmarks.
Does OrcaRouter store my data or use it for training?
OrcaRouter’s data handling policies are described in its terms of service. By default, the platform does not use customer data for model training. Data is processed in transit and may be cached for performance optimization. Review the OrcaRouter privacy policy for full details.
How do I call this model via an OpenAI-compatible API?
Use the base URL https://api.orcarouter.ai/v1 and set the model parameter to "qwen/qwen3-vl-235b-a22b-thinking". Authenticate with your OrcaRouter API key. The request format follows OpenAI’s chat completions API. For multimodal inputs, use a content array with type "text", "image_url", or "video_url".
Can I disable the thinking mode?
Yes. Pass the parameter "thinking": false in your API request. When disabled, the model returns only the final answer without the chain-of-thought reasoning. This reduces output token count and lowers cost.
What is the maximum output length?
The model can generate up to 40,960 tokens in a single response. This includes both the thinking chain (if enabled) and the final answer.
Is this model multilingual?
It is primarily optimized for English. Performance on non-English languages may be lower. The model may still handle some other languages, but for best results, use English prompts.
How does video input work?
Video input is provided as a URL to a video file. OrcaRouter samples frames from the video up to the context window limit. The model then treats the frames as a sequence, enabling reasoning about objects, actions, and temporal changes.

Embed this badge

Qwen: Qwen3 VL 235B A22B Thinking$0.40/M in4142ms p50via OrcaRouter
HTML <a href="https://www.orcarouter.ai/models/qwen/qwen3-vl-235b-a22b-thinking" target="_blank"> <img src="https://www.orcarouter.ai/embed/qwen/qwen3-vl-235b-a22b-thinking.svg" alt="Qwen: Qwen3 VL 235B A22B Thinking on OrcaRouter" /> </a>
Markdown [![Qwen: Qwen3 VL 235B A22B Thinking](https://www.orcarouter.ai/embed/qwen/qwen3-vl-235b-a22b-thinking.svg)](https://www.orcarouter.ai/models/qwen/qwen3-vl-235b-a22b-thinking)

Model card as data

GET /api/public/models/qwen/qwen3-vl-235b-a22b-thinkingOpen
Machine-readable:/llms.txt/llms-full.txt