GitHub radar
MiniCPM-V 4.5: GPT-4o-level vision, runs on a laptop
OpenBMB's 8.7B multimodal model analyzes images and high-FPS video locally via Ollama — 6 GB download, runs on a MacBook from 8 GB RAM.
MiniCPM-V 4.5 is a multimodal language model from OpenBMB (a research group at Tsinghua University), built on Qwen3-8B as its base. It understands single images, multiple images at once, and high-FPS video — all in one model without separate pipelines. Video compression works efficiently: six frames are packed into just 64 tokens instead of the typical 1,536, so long clips fit within the context without truncation. The model handles images up to 1.8 million pixels at any aspect ratio. On OCR benchmarks it outperforms GPT-4o-latest and Gemini 2.0 Pro. Context window is 40,000 tokens, and the download is 6.1 GB. Available via Ollama: `ollama run minicpm-v4.5`.
Why a vibe-coder should care
MiniCPM-V 4.5 lets you analyze images and video locally without Vision API costs or sending data to third-party servers. Screenshots, documents, photos, short clips — the model processes them on your own machine. With only 6 GB the download fits on any modern laptop, and performance beats much larger proprietary models on vision tasks.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install MiniCPM-V 4.5 by OpenBMB: ollama run minicpm-v4.5 — 6 GB download, runs on a MacBook with 8 GB RAM. After setup, show me how to send it an image or video frame and ask questions about what's in it.
A regular MacBook with 8 GB of RAM or more — the model takes 6 GB.
Open on Hugging Face▌ More finds