MiniCPM-V 4.7 lands on Ollama at 22GB

MiniCPM-V 4.7 marks the first time the MiniCPM-V series uses a mixture-of-experts design, where only a slice of the model runs per token. OpenBMB says it scored higher than Qwen3.6 35B-A3B on five image-understanding tests on average and in less time, though models with deliberate thinking modes still outscore it. The catch is that all 35B parameters still have to be held locally, so even though only about 3B do the work per word, a 4-bit copy is 21.2GB and an image component adds 1.12GB. A 1B version at 1.8GB is offered for those short on memory. The company's claim of 3GB memory and 13 tokens per second on an iPhone comes from a special llama.cpp build that preloads weights from storage, so ordinary Ollama will not match it. Japanese support and evaluation are not described, so organizations are advised to test with their own documents before relying on it.