MiniCPM locally (Ollama): multimodal compact

#Which MiniCPM version to install in 2026

MiniCPM-V is OpenBMB’s compact vision-language model series, developed by THUNLP (Tsinghua University) and ModelBest, with weights and code released under the Apache 2.0 license. In Ollama, three tags coexist and cause confusion. The minicpm-v tag, without a suffix, still corresponds to MiniCPM-V 2.6, an 8-billion-parameter model weighing 5.5 GB with 32K context. The minicpm-v4.5 tag loads a newer 8-billion-parameter version (6.1 GB, 40K context). The minicpm-v4.6 tag, added to the official Ollama library on June 25, 2026, according to the project repository, weighs 1.6 GB and has approximately 1.3 billion parameters. To explore local vision on a small machine, start with 4.6; keep 4.5 for comparison on dense documents, because nothing guarantees that a 1.3-billion-parameter model will read tightly packed text as well as an 8-billion-parameter model.

#What the series delivers, and what to put into perspective

The series aims for the best capacity-to-size ratio. According to the OpenBMB repository, MiniCPM-V 4.5 scores 77.0 on OpenCompass and, with 8 billion parameters, surpasses GPT-4o-latest, Gemini 2.0 Pro, and Qwen2.5-VL 72B. For version 4.6, the project reports a visual encoder whose compute requirements drop by more than 50%, optional visual-token compression (4x or 16x), and a token throughput approximately 1.5 times that of Qwen3.5-0.8B. The Ollama sheet for the same model reports 2.4 times: both figures come from the project, and neither page details a measurement protocol. These are claims, not independent measurements: benchmark them on your hardware.

Compact multimodal

Text and images, multiple images at once, and video. According to the repository, 4.5 compresses up to 6 video frames into 64 tokens, allowing it to process more frames without additional overhead for the language model.

High-resolution OCR

The 4.5 handles images of any aspect ratio up to 1.8 million pixels (for example, 1344 x 1344). The publisher claims it leads on OCRBench and PDF analysis (OmniDocBench) among general-purpose multimodal models; the 4.6 is said to reach Qwen3.5 2B-level performance on several benchmarks, including OCRBench.

Permissive license

The repository announces weights and code under the Apache 2.0 license, and the Hugging Face cards for versions 4.5 and 4.6 carry the same notice. Check the exact license for the specific version you download.

Many formats

The project offers GGUF, BNB, AWQ, and GPTQ variants, and advertises compatibility with SGLang, vLLM, llama.cpp, and Ollama.

#How much memory, really

These figures describe the weights and basic execution; context and images require additional memory. Images are converted into visual tokens that occupy the context window, and Ollama’s documentation specifies that it applies 4,000 context tokens by default under 24 GiB of VRAM; increasing the value increases the required memory. The advertised 256K context for 4.6 is therefore not allocated by default. Finally, the repository table indicates CPU execution for the GGUF version of 4.6 without publishing a speed: measure it before promising real-time performance.

#Install MiniCPM-V with Ollama

Ollama handles the download, visual projector, and API server: there’s nothing else to install. Start by updating Ollama, because 4.6 is recent in the library. The minicpm-v tag page already mentions that it required Ollama 0.3.10 or later.

1. 01
   Download the model

   ollama pull minicpm-v4.6 récupère les poids du modèle de langage et le projecteur visuel, soit environ 1,6 Go. Pour le 4.5, utilisez minicpm-v4.5 (6,1 Go).
2. 02
   Start a first chat

   ollama run minicpm-v4.6 ouvre une session interactive. Posez une question texte avant de tester la vision.
3. 03
   Send an image

   Ollama's documentation shows the file path placed before the question, for example ollama run minicpm-v4.6 ./photo.jpg followed by “Describe this image.”
4. 04
   Connect an interface

   Once pulled, the model appears in Open WebUI or any client that points to port 11434.

# Le plus léger (1,6 Go)
ollama pull minicpm-v4.6

# La version 8 milliards, pour comparer
ollama pull minicpm-v4.5

# Vérifier ce qui est installé, puis discuter
ollama list
ollama run minicpm-v4.6 ./photo.jpg Décris cette image

#Vision and OCR for everyday use

MiniCPM-V is well suited to document tasks: reading an invoice, transcribing a screenshot, extracting table rows, and describing a photo. High-resolution images matter a lot for OCR, because fine text that is unreadable in a thumbnail becomes usable at full resolution. Write precise prompts: “Describe this image” produces a general description, whereas “Faithfully transcribe all visible text, preserving the layout” produces a much cleaner result. To extract data, request an explicit format: JSON, Markdown, or a table.

import base64, requests

with open("facture.png", "rb") as f:
    img = base64.b64encode(f.read()).decode()

r = requests.post("http://localhost:11434/api/generate", json={
    "model": "minicpm-v4.6",
    "prompt": "Extrais le total, la date et le numéro de facture en JSON.",
    "images": [img],
    "stream": False,
})
print(r.json()["response"])

Ollama's REST API expects the image encoded as base64 in the images field, as its documentation notes; the SDKs also accept a file path. If you use Transformers instead of Ollama, the repository documents a useful setting for 4.6: downsample_mode defaults to 16x, while 4x preserves four times as many visual tokens for finer detail. This is the setting to try when very small text is misread.

When should you prefer a dedicated OCR engine? If you process thousands of pages and traceability is paramount, a specialized engine produces visible errors instead of plausible values. PaddleOCR-VL, a model with fewer than one billion parameters, scores 96.33% on OmniDocBench v1.6 according to its publisher, while Tesseract remains the lightweight choice for clean text. MiniCPM-V retains the advantage when you also want to describe the image or answer questions about it.

• PaddleOCR: OCR that understands the page
• Tesseract OCR: read a scan locally
• Extract invoice data end to end

#Compared with Qwen3-VL and other compact models

Qwen3-VL is MiniCPM-V's direct competitor in the compact vision-model category, available in Ollama, notably in 2-, 4-, and 8-billion-parameter versions. Both cover the same ground: image description, OCR, and questions about a document. OpenBMB's README claims that MiniCPM-V 4.6 outperforms the larger Gemma4-E2B-it model and is more efficient than Qwen3.5-0.8B: these are comparisons published by the vendor using its own evaluations.

The choice depends more on your images than on a spec sheet. Both families run under the same Ollama and use the same API: switching between them means changing the model name in the request. You can keep both and route requests based on the task.

ollama run minicpm-v4.6 ./ticket.jpg Transcris le texte de ce ticket
ollama run qwen3-vl:2b ./ticket.jpg Transcris le texte de ce ticket

• Qwen3-VL locally: the reference vision model
• Moondream: the vision model that runs anywhere

#Edge, mobile, and server: what the project documents

MiniCPM was designed for devices. The repository states that MiniCPM-V 4.6 runs on iOS, Android, and HarmonyOS, with open adaptation code, and shows raw screen recordings on an iPhone 17 Pro Max, a Redmi K70, and a HUAWEI nova 14. An application repository offers the download and deployment guide. To serve multiple users, the project announces compatibility with vLLM and SGLang, in addition to llama.cpp and Ollama.

1. 01
   Local document digitization

   A mini PC or NAS with an entry-level GPU turns a folder of scans into searchable text without sending a sensitive document to the cloud.
2. 02
   Offline assistant

   On a Apple Silicon laptop or a modest card, 4.6 reads the screen and responds to screenshots, even without a connection.
3. 03
   Pipeline preprocessing

   Ahead of a heavier system, it sorts by document type and obvious fields. Only the difficult cases are sent to a large model.
4. 04
   Bulk alt text

   Generate image descriptions for a website or media library in batches, with no per-call cost once the hardware has paid for itself.

• SGLang: serving a local LLM to multiple users
• vLLM: what is it, who is it for, and when should you use it?

#Troubleshooting

The model ignores the image

Check the file path and ensure it is accessible from where Ollama is running. In the API, the image must be base64-encoded in the images field.

Sloppy or truncated OCR

The image is probably too compressed or too low-resolution. Provide a sharp, larger version and explicitly request a faithful transcription with the layout preserved. With Transformers, try downsample_mode 4x.

Slow responses

On pure CPU, this is expected. On GPU, check with ollama ps that the model has not been partially offloaded to RAM because of an overly large context or image.

Unstructured output

For reliable JSON, enforce the schema in the prompt and, if your client supports it, use Ollama’s structured format.

Unexpected version

The minicpm-v tag without a suffix still points to 2.6. Explicitly name minicpm-v4.6 or minicpm-v4.5 for reproducible behavior.

#Go further

• Install Ollama in 5 minutes
• A local multimodal vision LLM with Ollama
• Choose your quantization
• Open WebUI with Ollama
• Source: official MiniCPM-V repository from OpenBMB
• Source: Ollama page from minicpm-v4.6
• Source: MiniCPM-V 4.6 Hugging Face model card
• Source: Ollama documentation on vision