MiniCPM-V 4.5: specs, benchmarks, and how to run it on Lyceum

Get started: call MiniCPM-V 4.5 on Lyceum

You can integrate MiniCPM-V 4.5 into your application using Lyceum's OpenAI-compatible API. Because our inference API is designed as a drop-in replacement for standard OpenAI SDKs, you only need to update your base URL and API key to start processing images, documents, and video frames. You can avoid writing custom integration code or managing complex multimodal container deployments.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="openbmb/MiniCPM-V-4_5",
 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 MiniCPM-V 4.5

MiniCPM-V 4.5 is available on Lyceum's serverless inference API, which is optimized for cost-efficient, high-throughput workloads like batch document processing and continuous video stream analysis. The model is priced at $0.66 per million input tokens and $1.11 per million output tokens, and Lyceum's catalogue lists it with a 128K context window. All inference requests for this model are processed exclusively in our eu-north1 region. That means GDPR-compliant processing in European data centres, so sensitive visual data is not routed through US-based servers. Prompts and outputs are not retained after processing.

What MiniCPM-V 4.5 is good at

High-FPS and long video understanding

Most multimodal large language models struggle with video inputs because the token count scales linearly with the number of frames. This rapidly leads to out-of-memory errors, truncated context windows, and prohibitively high inference costs. MiniCPM-V 4.5 solves this fundamental bottleneck using a unified 3D-Resampler architecture. It compresses six 448x448 video frames into just 64 tokens, an unprecedented 96x compression rate. This allows the model to process high-FPS video (up to 10 FPS) and long-duration videos efficiently, achieving state-of-the-art results on benchmarks like Video-MME and LVBench for models under 30B parameters.

High-resolution OCR and document parsing

Built on the advanced LLaVA-UHD architecture, MiniCPM-V 4.5 can process high-resolution images of any aspect ratio, supporting up to 1.8 million pixels (for example, a 1344x1344 image). Crucially, it achieves this while using 4x fewer visual tokens than standard MLLMs. This makes the model exceptionally strong at Optical Character Recognition (OCR), extracting dense text from scanned documents, and parsing complex tables into structured Markdown formats for downstream data processing pipelines.

Controllable hybrid fast/deep thinking

The model introduces a novel hybrid reasoning approach. It supports a "fast thinking" mode for standard, latency-sensitive visual queries, and a "deep thinking" mode for complex problem-solving and multi-step reasoning tasks. This allows machine learning engineers to toggle between raw speed and analytical depth depending on the specific requirements of their production workload, optimizing both latency and compute costs.

Limitations and trade-offs

Where MiniCPM-V 4.5 falls short

While MiniCPM-V 4.5 performs competitively for its size in multimodal tasks, it is fundamentally still an 8-billion parameter model. For highly nuanced, zero-shot reasoning on extremely complex visual logic puzzles or deep semantic understanding of abstract imagery, the largest proprietary frontier models may still hold an edge in raw reasoning depth, despite MiniCPM-V 4.5's highly competitive benchmark scores.

Additionally, the model's "deep thinking" mode, while useful for complex reasoning, can sometimes interfere with its grounding and spatial localization capabilities. Machine learning engineers testing the model have noted that enabling the thinking mode can occasionally degrade its ability to output precise bounding boxes or exact spatial coordinates. For strict localization tasks, such as identifying the exact pixel coordinates of a defect in a manufacturing pipeline, the fast mode is generally preferable, or a dedicated object detection model might be required.

Finally, while the 96x video token compression is a massive efficiency gain for long-video understanding, it inherently involves some loss of granular, frame-by-frame detail. For specialized use cases requiring pixel-perfect tracking of microscopic changes across consecutive frames, such as certain types of medical imaging analysis or high-speed anomaly detection, the heavy compression might obscure subtle visual shifts compared to processing raw, uncompressed frames through a dedicated vision transformer.

Benchmarks and how it compares

MiniCPM-V 4.5 benchmark results

MiniCPM-V 4.5 was evaluated across a wide range of multimodal benchmarks, demonstrating performance that rivals much larger models. It achieves an average score of 77.0 on the OpenCompass multimodal evaluation suite, making it one of the most capable vision-language models under 30B parameters.

Source: OpenBMB MiniCPM-V 4.5 model card, read 3 August 2026.

Comparison to sibling models

Compared to its predecessor, MiniCPM-V 2.6, the 4.5 version introduces the 3D-Resampler for massive video token compression and the hybrid thinking mode, resulting in significantly better performance on long-video benchmarks like Video-MME. When compared to other vision-language models such as Qwen2.5-VL 7B, MiniCPM-V 4.5 offers superior video compression. According to the technical report, MiniCPM-V 4.5 achieves state-of-the-art performance on Video-MME while using just 46.7% of the GPU memory cost and 8.7% of the inference time compared to Qwen2.5-VL 7B. This makes it the clear choice for teams processing large volumes of video data where compute efficiency is paramount. A multimodal model built for the same job is covered in our notes on Nemotron-3-Nano-Omni and how to run it.

Using it in production

Production configuration for MiniCPM-V 4.5

When deploying MiniCPM-V 4.5 for production workloads, understanding its token economics is critical for optimizing costs. Because of its LLaVA-UHD architecture, high-resolution images (up to 1.8M pixels) are processed using significantly fewer visual tokens than traditional vision models. For video, the 96x compression rate means you can pass substantial video context without exhausting the context window or inflating your per-token costs.

At $0.66 per million input tokens and $1.11 per million output tokens on Lyceum's serverless inference API, the model is highly cost-effective for batch OCR and video analysis. For example, processing a batch of 1,000 document images, assuming roughly 1,000 input tokens per image and 200 output tokens for the extracted Markdown, would cost approximately $0.66 for the input and $0.22 for the output, totaling less than $1.00 for the entire batch. This makes it highly viable for large-scale enterprise document parsing.

For video workloads, you should extract frames at your desired FPS (up to 10 FPS is supported efficiently) and pass them as a sequence of image inputs in the API call. The model's architecture will automatically handle the compression. Ensure your application handles standard OpenAI SDK streaming responses if you are utilizing the model's text generation capabilities for real-time video narration or chat. You can also toggle the hybrid thinking mode via system prompts depending on whether you need low-latency responses or deep reasoning.

Running MiniCPM-V 4.5 on EU-sovereign infrastructure

Why run MiniCPM-V 4.5 on Lyceum

For European AI startups and enterprise engineering teams, data residency is often a strict, non-negotiable requirement. Processing sensitive documents, medical imagery, or proprietary factory video feeds through US-based hyperscalers can violate internal compliance policies or GDPR mandates. Lyceum solves this by serving MiniCPM-V 4.5 from our eu-north1 region, so inference for this model runs in European data centres.

By using Lyceum, you get the ease of an OpenAI-compatible API without the compliance risks of routing data outside the EU. Lyceum serves this model from European data centres in Spain, Paris and the Nordics. Serverless Inference is billed per token with no base fee and no service tiers, and GPU VMs are billed per second, so you pay only for what you process.

Furthermore, Lyceum's open-stack transparency, powered by vLLM and NVIDIA Dynamo, ensures you avoid vendor lock-in. You can smoothly transition from our Serverless Inference API to a dedicated self-hosted LLM API on EU infrastructure as your workload scales. You can start prototyping with MiniCPM-V 4.5 on a pay-per-token basis today and move to reserved VMs later, maintaining the exact same codebase and compliance posture throughout your entire scaling journey.

Sources

[1] MiniCPM-V 4.5 Hugging Face Model Card; [2] MiniCPM-V GitHub Repository; [3] MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe (arXiv); [4] MiniCPM-V 4.5 on Ollama; [5] MiniCPM-V 4.5 on ModelScope