Qwen3-VL 235B-A22B Instruct — open-weight vision-language model, 235B total / 22B active params, 256k context, no thinking mode.
Qwen3 VL 235B A22B Instruct is a vision-language variant of the Qwen3 model series, built by Alibaba Cloud. It uses a mixture-of-experts design with 235 billion total parameters, of which…
The model excels at tasks where images and text interact. It can answer questions about the content of an image, describe scenes in detail, read text from documents or signs, and reason about relationships between objects in a picture. It also handles multiple images in a single conversation, allowing comparative analysis or sequential reasoning. The instruction tuning enables the model to follow complex multimodal prompts, such as 'Explain why this chart shows a trend' or 'Extract all numerical values from this table.' These capabilities stem from the large MoE backbone and specialized vision-language training.
Being an Instruct variant, the model has been fine-tuned to follow user instructions with high fidelity across both text-only and multimodal prompts. It can handle multi-turn dialogues where images are introduced incrementally, maintain context over several exchanges, and follow formatting instructions (e.g., output as JSON, bullet points, or step-by-step). It performs well on tasks that require adherence to constraints, such as 'Answer only if the image contains a dog.' This makes it suitable for applications where consistent, rule-following behavior is important.
For text-only tasks with no visual component, the 235B MoE model may be overkill; smaller dense models or other text-only Qwen variants would be more cost-effective. Similarly, if your images are simple (e.g., basic logos, single objects) or you need extremely high throughput on a budget, a smaller vision model like Qwen2-VL 7B could suffice. This model is most justified when you need to handle complex, high-resolution images, documents with dense text, or tasks requiring deep multimodal reasoning. On OrcaRouter, you can compare pricing and switch model IDs easily.
No. Qwen3 VL 235B A22B Instruct is a text-generation model that accepts images as input only. It does not generate images, audio, or any other modality. Output is always a text string. If you need image generation, you would use a separate model. This model's strength lies in understanding and reasoning about visual input. It can, for example, describe what an image looks like or answer questions about it, but it cannot create, edit, or transform images itself.
The reported MMLU-Pro score for this model is 82.3. MMLU-Pro is an enhanced version of the Massive Multitask Language Understanding benchmark, covering 57 subjects across STEM, humanities, and social sciences. A score of 82.3 indicates strong general knowledge and reasoning across diverse topics. It is a single-number aggregate; performance may vary per subject. This score places the model among the top performers in its size class, especially considering the MoE architecture that activates only a fraction of its total parameters per query.
Strengths include high accuracy on multimodal reasoning benchmarks, strong instruction following, and cost efficiency relative to dense models of similar total parameter count. The MoE design allows it to scale capacity without proportional compute cost. Limitations include higher absolute cost compared to smaller models, potential for increased latency due to the large number of total parameters (even though only 22B are active, the full model must be loaded in memory). It may also occasionally hallucinate fine details in images or fail on highly ambiguous visual inputs. Performance on specific domain tasks (e.g., medical imaging) is not guaranteed.
Inference speed depends on the hardware backend used by OrcaRouter and the current load. As an MoE model with 235 billion total parameters, it requires significant GPU memory even though only 22 billion parameters are activated per token. This typically results in higher per-request latency compared to smaller dense models (e.g., 7B-70B parameters). Throughput is also influenced by batch size and sequence length. For latency-sensitive applications, consider using a smaller model or optimizing for shorter prompts. OrcaRouter does not provide specific latency guarantees but aims for production-grade responsiveness.
OrcaRouter charges exactly the provider's listed rate: $0.40 per 1 million input tokens and $1.60 per 1 million output tokens. There is no additional markup. Both input and output tokens are counted according to OpenAI's tokenization rules, which may differ slightly from the model's internal tokenizer. Images are also billed as tokens based on their dimensions and detail level, following the provider's policy. You can estimate costs by multiplying your expected token usage by these rates.
The cost per token is higher than smaller or dense models, but the MoE architecture provides more capacity per active parameter than a dense model of equivalent active size. For complex multimodal tasks, this model may complete the task in fewer tokens or with higher accuracy, potentially reducing total spend. However, for simple tasks, a cheaper model would lower costs. On OrcaRouter, you can switch models on the fly, allowing you to use this model only when needed. There are no minimum monthly commitments or hidden fees.
OrcaRouter does not currently disclose specific caching policies for this model. Token-level caching may be applied by the underlying provider but is not guaranteed. There are no bulk discounts or tiered pricing mentioned; rates are flat per token. For high-volume users, it may be worth contacting OrcaRouter support for potential custom arrangements. In general, pricing is transparent and usage-based.
Images are converted into a token count based on their resolution and detail setting. The exact formula is defined by the provider (Qwen) and may vary. Typically, higher resolution images consume more tokens. OrcaRouter passes through the provider's tokenization without alteration. You can control costs by resizing images or using the low-detail mode if the model supports it. Always refer to the provider's documentation for the precise image token calculation. Input tokens for images are counted toward the $0.40 per million rate.
You send a POST request to https://api.orcarouter.ai/v1/chat/completions with an OpenAI-compatible JSON payload. Set the model field to "qwen/qwen3-vl-235b-a22b-instruct". Include a messages array where each message can contain text and/or image content. For images, use the standard OpenAI image content format (URL or base64). Authentication is via a Bearer token (your API key). Example parameters: temperature, max_tokens, top_p, and stop sequences work identically to the OpenAI API. OrcaRouter's documentation provides full detail.
All standard chat completions parameters are supported: temperature (0-2, default 1), top_p (0-1, default 1), max_tokens (number of output tokens), stop (list of strings), presence_penalty, frequency_penalty, and seed for reproducibility. The model also respects system messages for setting behavior. For multimodal requests, you include an image part in the content of a user message. There are no model-specific parameters exposed; the API is kept generic for compatibility. If you need to control the MoE routing, that is handled internally.
Yes. Because OrcaRouter uses the OpenAI API format, migration is straightforward. Replace your existing base URL and model ID with OrcaRouter's endpoint and "qwen/qwen3-vl-235b-a22b-instruct". Ensure your API key is valid and that you have sufficient credits. The message format for images is identical to OpenAI's (using the 'image_url' content type). You may need to adjust token count estimates due to different tokenization. No changes to your application's logic are required beyond endpoint and model name.
Qwen3 VL 235B A22B Instruct and GPT-4V both handle multimodal inputs. Key differences: Qwen3 VL uses MoE with 22B active parameters, while GPT-4V is a proprietary dense model of an undisclosed size. Qwen3 VL pricing through OrcaRouter is $0.40/$1.60 per million tokens, which is significantly lower than typical GPT-4V rates. Benchmark scores are not directly comparable because MMLU-Pro and other tests use different subsets. GPT-4V has broader ecosystem support, but Qwen3 VL offers a cost advantage for high-volume multimodal workloads on OrcaRouter.
Claude 3.5 Sonnet is a dense model from Anthropic with strong text and vision capabilities. It does not use MoE, so its total capacity is less than Qwen3 VL's total parameters but its active capacity per inference is higher than 22B. Pricing for Claude 3.5 Sonnet is generally higher per token (approximately $3.00/$15.00 per million tokens). Qwen3 VL offers a much lower cost for multimodal tasks. However, Claude models have a larger context window (200K tokens). Choice depends on budget, context length needs, and preferred provider.
OrcaRouter likely offers other Qwen models such as Qwen2.5 7B, Qwen2.5 72B, or Qwen2-VL variants. The Qwen3 VL 235B A22B Instruct is the largest multimodal MoE model in the Qwen lineup. Compared to Qwen2-VL 72B, it has more total parameters and an MoE architecture, which can yield better performance on complex tasks while maintaining reasonable inference cost. It is more expensive per token than smaller Qwen models but may be cost-effective if it reduces the need for multiple calls or high token consumption.
OpenAI-compatible — keep the SDK you already use
https://api.orcarouter.ai/v1from openai import OpenAI
client = OpenAI(
base_url="https://api.orcarouter.ai/v1",
api_key="$ORCAROUTER_API_KEY",
)
response = client.chat.completions.create(
model="qwen/qwen3-vl-235b-a22b-instruct",
messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)| Input / 1M tokens | $0.400 |
| Output / 1M tokens | $1.60 |
| Currency | USD |
Estimate based on list price
Estimate only — actual token counts depend on the provider's tokenizer.
GET /api/public/models/qwen/qwen3-vl-235b-a22b-instructOpen @misc{orcarouter_qwen3_vl_235b_a22b_instruct,
title = {Qwen3 VL 235B A22B Instruct API},
author = {Qwen},
year = {2025},
howpublished = {OrcaRouter},
url = {https://www.orcarouter.ai/models/qwen/qwen3-vl-235b-a22b-instruct}
}Qwen. (2025). Qwen3 VL 235B A22B Instruct API. OrcaRouter. https://www.orcarouter.ai/models/qwen/qwen3-vl-235b-a22b-instruct