Qwen2.5-VL-72B: specs, benchmarks, and how to run it on Lyceum

Get started: call Qwen2.5-VL-72B on Lyceum

Integrate with the OpenAI SDK

Integrate Qwen2.5-VL-72B into your application without learning new frameworks or writing custom API wrappers. The platform provides a fully managed, OpenAI-compatible Serverless Inference API. Updating the base URL and API key routes multimodal requests directly to European data centers.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<base URL from your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="Qwen/Qwen2.5-VL-72B-Instruct",
 messages=[{"role": "user", "content": [
 {"type": "text", "text": "What's in this image?"},
 {"type": "image_url", "image_url": {"url": "<image-url>"}},
 ]}],
 )
 print(response.choices[0].message.content)

Pricing and region for Qwen2.5-VL-72B

Route inference traffic to Lyceum to pay strictly for the compute consumed. Qwen2.5-VL-72B is served through a single Serverless Inference endpoint billed per token, with no service tiers to choose between. It is optimized for high-capability workloads requiring precision and reasoning depth, and carries a 128,000-token context window. The model is hosted in the eu-north1 region, ensuring that images, videos, and text prompts never leave European jurisdiction.

The pricing for Qwen2.5-VL-72B is competitive for a 72-billion parameter multimodal model. You are billed at $0.25 per million input tokens and $0.75 per million output tokens. Lyceum serves the model from European data centres in Spain, Paris and the Nordics, and offers these rates without minimum commitments, base fees, or egress charges. You scale from zero and pay only for the tokens processed.

What Qwen2.5-VL-72B is good at

Advanced document parsing and structured output

Qwen2.5-VL-72B excels at extracting structured data from complex, text-heavy images. The model was trained on a massive corpus of document data, enabling it to parse invoices, forms, tables, charts, and even mathematical equations or chemical formulas. It can output stable JSON representations of this data, making it an ideal choice for automated data entry, financial document processing, and enterprise OCR pipelines.

Spatial localization and visual agent capabilities

Unlike earlier vision-language models that only describe images globally, Qwen2.5-VL-72B possesses precise spatial localization capabilities. It can accurately pinpoint objects within an image and output their exact coordinates using bounding boxes or points. This spatial awareness allows the model to act as a visual agent. It can reason about user interfaces, dynamically direct tools, and execute tasks in real-world scenarios, such as operating computer desktops or mobile device screens based on visual feedback.

Long-video comprehension

Qwen2.5-VL-72B introduces a leap in temporal understanding. By implementing dynamic FPS sampling and absolute time encoding, the model can process videos extending over an hour in length. It natively perceives temporal dynamics without relying on traditional normalization techniques. This allows the model to not only summarize long videos but also pinpoint specific events down to the second, making it highly effective for security footage analysis, meeting transcription, and automated video editing workflows.

Limitations and trade-offs

Where Qwen2.5-VL-72B falls short

Qwen2.5-VL-72B is a capable multimodal model, but its 72-billion parameter scale introduces infrastructure challenges for teams attempting to self-host. Running inference on a model of this size, especially with large image or video contexts, requires multiple high-end GPUs. Serving the model in BF16 precision typically demands at least two to four NVIDIA H100 or A100 GPUs to fit the weights and KV cache into VRAM. Dedicated deployments are expensive for workloads with low or bursty traffic.

While the model is designed to be agentic, integrating its tool-calling capabilities into standard inference pipelines can be complex. Developers using open-source serving frameworks like vLLM have reported that Qwen2.5-VL-72B does not always support tool usage out of the box without custom chat templates. Merging the multimodal input format with the tool-call syntax requires prompt engineering and framework configuration.

Processing high-resolution images and long videos consumes context window capacity rapidly. Although Qwen2.5-VL-72B utilizes a dynamic-resolution Vision Transformer to reduce computational overhead, feeding hours of video or dozens of high-megapixel images will exhaust the context limit and increase time-to-first-token. Teams must manage image resolution and video frame sampling rates to balance reasoning quality against latency and token costs.

Benchmarks and how it compares

Qwen2.5-VL-72B benchmark results

Qwen2.5-VL-72B was evaluated against both open-source and proprietary multimodal models. The official technical report states the 72B flagship model matches or exceeds the performance of proprietary models like GPT-4o and Claude 3.5 Sonnet in document parsing and spatial reasoning tasks.

Source: Qwen2.5-VL Technical Report (arXiv), Tables 3 and 5, read 3 August 2026.

Compared to its sibling models, Qwen2.5-VL-72B provides an increase in reasoning depth. While the smaller Qwen2.5-VL-7B is efficient for basic image captioning and the Qwen2.5-VL-32B offers a middle ground for edge deployments, the 72B variant is required for complex agentic workflows, dense document omni-parsing, and hour-long video comprehension. On OCRBench_v2 the 72B model scores 61.5 on the English track and 63.7 on the Chinese track, which the Qwen2.5-VL technical report (Table 5, read 3 August 2026) describes as exceeding the best comparison model, Gemini 1.5-Pro, by 9.6 percent and 20.6 percent respectively. For enterprise applications where accuracy is paramount, the 72B model is the choice in the Qwen lineup.

Using it in production

Production configuration for Qwen2.5-VL-72B

Deploying Qwen2.5-VL-72B in production requires careful management of multimodal inputs to optimize both latency and cost. The model's published configuration sets a 128,000-token maximum position embedding, allowing you to pass multiple high-resolution images or long video sequences in a single API request. However, because image tokens are calculated based on the dynamic resolution processing of the Vision Transformer, sending raw, uncompressed 4K images will rapidly inflate your input token count. We recommend resizing images to the minimum resolution necessary for your specific task, such as 1024x1024 for standard OCR, to keep inference fast and cost-effective.

When calling the model via Lyceum Technology, you are using Serverless Inference. It runs on high-end NVIDIA GPUs in our eu-north1 region, providing the compute density required for a 72-billion parameter model, and it is billed per token with no service tiers to select. That suits complex reasoning, structured JSON extraction, and agentic tool use. Retrieval pipelines need a different kind of model, which we cover in our notes on Qwen3-Embedding-8B and how to run it.

Lyceum charges strictly per token, keeping costs predictable. At $0.25 per million input tokens and $0.75 per million output tokens, a typical document parsing workload is economical. For example, processing 1,000 invoices, where each invoice consumes roughly 2,000 input tokens for the image and prompt and generates 300 output tokens for the extracted JSON, results in a total consumption of 2 million input tokens and 300,000 output tokens. This batch costs $0.50 for input and $0.225 for output, totaling $0.725. You achieve OCR and data extraction without the overhead of maintaining dedicated infrastructure.

Running Qwen2.5-VL-72B on EU-sovereign infrastructure

Why run Qwen2.5-VL-72B on Lyceum

For European enterprises and AI startups, data residency is not optional. When processing sensitive multimodal data, such as medical scans, financial documents, or factory floor security footage, sending that data to US-based hyperscalers or API providers introduces severe compliance risks. Lyceum Technology solves this by serving Qwen2.5-VL-72B entirely from our eu-north1 region. Your images, videos, and prompts are processed on EU-sovereign infrastructure, which means GDPR-compliant processing in European data centres.

Lyceum serves Qwen2.5-VL-72B from European data centres in Spain, Paris and the Nordics. Pricing is per token with no base fees and no egress charges. You get the performance of a 72-billion parameter model without the financial burden of provisioning dedicated H100 clusters.

The platform prevents vendor lock-in through open-stack transparency. The inference engine is built on open-source technologies like vLLM and NVIDIA Dynamo, ensuring that you can self-host your LLM API on EU infrastructure if scale demands it. A drop-in OpenAI-compatible API makes it trivial for engineering teams to switch from proprietary US models to Qwen2.5-VL-72B. Updating the base URL and API key instantly upgrades the application with multimodal capabilities while securing data within Europe.

Sources

[1] Qwen2.5-VL Technical Report; [2] Qwen2.5-VL Official Blog; [3] Qwen2.5-VL-72B-Instruct on Hugging Face; [4] Qwen2.5-VL GitHub Repository; [5] Qwen2.5-VL-72B-Instruct config.json (read 3 August 2026)