MiniCPM-V 4.7 on a Mac: memory, speed and Ollama

Q: Is MiniCPM-V 4.7 on Ollama?

Yes, as openbmb/minicpm-v4.7, not in the main library.

Short answer

MiniCPM-V 4.7 35B-A3B is a mixture-of-experts vision model that uses about 3B parameters per token. With the standard files it needs a 32 GB Mac or more: the Q4_K_M plus vision projector is about 22.3 GB, and Ollama’s openbmb/minicpm-v4.7 tag is 22 GB. With OpenBMB’s split “LES” bundle, one tester ran it on a 16 GB M4 Mac mini at about 8 tokens a second using 3.5 GB of RAM. The 1B model fits anything.

Sizes by Mac ↓Published Mac runs ↓Model-fit calculator →

What MiniCPM-V 4.7 is

OpenBMB published MiniCPM-V 4.7 in early October 2026 in two sizes: a 35B-A3B mixture-of-experts model, the first MoE in the series, and a dense 1B. Both read images and video as well as text, and both are Apache 2.0. The language part is Qwen 3.5, and the vision encoder is SigLIP2. OpenBMB says the 35B-A3B beats Qwen 3.6 35B-A3B in instruct mode while taking less time. That is their claim, not a test I have seen repeated.

The “on-device” part is a technique OpenBMB calls Lookahead Expert Scheduling (LES). The experts stay on the SSD, and the runtime loads the ones it expects to need next. That is how they get to 13+ tokens a second with 3 GB of memory on an iPhone 17 Pro Max (vendor figure). They also ship DSpark heads for speculative decoding.

MiniCPM-V 4.7 file sizes and the Mac each one needs

MiniCPM-V 4.7 downloads (vendor file sizes unless marked)
Build	Size, GB	Mac I would plan for
35B-A3B Q8_0 GGUF	36.9	64 GB
35B-A3B Q4_K_M GGUF	21.17	32 to 48 GB
35B-A3B AWQ int4 (for vLLM)	22.7	Not a Mac path
35B-A3B mix_q3gud_im GGUF	15.5	24 GB, with little else open
35B-A3B IQ3_XS GGUF	14.48	24 GB
35B-A3B LES bundle (core + 40 expert files + projector)	16.62 on disk	16 GB, slower
Vision projector (mmproj)	1.12	Add to any row above
1B Q4_K_M GGUF	0.67	Any Mac

The LES bundle size comes from lilting.ch’s byte count (community). Everything else is from OpenBMB’s file lists. Note that a mixture-of-experts model still needs room for all of its experts unless the runtime streams them from disk, so the 3B active figure does not shrink the download.

Published runs on 16, 24 and 48 GB Macs

MiniCPM-V 4.7 35B-A3B on Apple silicon (community and secondary)
Machine	Setup	Speed	Memory	Source
M4 Mac mini, 16 GB	llama.cpp-omni, LES bundle	7.6–8.0 tok/s text; 5.8–6.0 with images	3.5 GB peak (text), 4.6 GB (images)	lilting.ch, community
Mac, 24 GB	llama-server, mix_q3gud_im	About 25 tok/s resident; up to 2 min a request when paging	15.5 GB weights	Micdrop, community
M4 Pro, 48 GB	oMLX, 4-bit (oQ4e)	75–82 tok/s	About 21 GB peak	oMLX via LLMCheck, secondary
M4 Max	oMLX, MXFP4	About 87 tok/s	Not reported	oMLX via LLMCheck, secondary

On the 16 GB Mac mini, lilting.ch ran three test sets. The model scored full marks on fact checks and event counting, as the larger hosted models did, and 23 of 24 on a Python interval-merging problem. It was slow, at about 93 seconds a coding answer, and it got arithmetic wrong with thinking mode off. Turning on DSpark added about 9.5% speed. They had to use OpenBMB’s fork, llama.cpp-omni, because the LES bundle does not run in mainline llama.cpp.

On a 24 GB Mac, Micdrop found the 35B-A3B was the first local model that handled every tool call in their voice-agent tests, in French as well as English. But with an editor and a browser open, macOS paged the weights out, and one request took up to two minutes. They kept Qwen3 4B as their default on that machine.

How to run MiniCPM-V 4.7 on a Mac

Ollama. ollama run openbmb/minicpm-v4.7 pulls the 35B-A3B at Q4_K_M (22 GB). The 1B tag is 1.8 GB. The model lives in OpenBMB’s namespace; library/minicpm-v4.7 returned 404 on 4 November. Micdrop reports that Ollama 0.32.15 rejected the mix_q3gud_im file, so on a 24 GB Mac the smallest tag that worked for them was Q4_K_S at 19.9 GB.
llama.cpp. Mainline support arrived in PR 29416 (10 October). Pass the projector with --mmproj for images.
MLX. mlx-vlm added the model in 0.7.8, and LLMCheck recommends 0.7.9 or later for quantized builds.
16 GB Macs. Use the LES bundle with OpenBMB’s llama.cpp-omni, and expect single-digit speed.

Ollama: the 35B-A3B at Q4_K_M from OpenBMB’s namespace

ollama run openbmb/minicpm-v4.7

If you only need OCR

OpenBMB released MiniCPM-OCR 1.0 on 28 October. It is based on MiniCPM-V 4.6, also Apache 2.0, and the Q4_K_M GGUF is 0.53 GB. For reading invoices and scanned pages, that is a much smaller download than the 35B-A3B. I have not compared their accuracy.

FAQ: MiniCPM-V 4.7 on a Mac

Can MiniCPM-V 4.7 run on a 16 GB Mac?

The 35B-A3B can, with OpenBMB’s LES bundle and llama.cpp-omni. One test on a 16 GB M4 Mac mini measured 7.6 to 8.0 tokens a second with 3.5 GB of RAM. The standard 21 GB file does not fit.

How much memory does MiniCPM-V 4.7 need?

About 22.3 GB for the Q4_K_M file and projector, or 22 GB for the Ollama tag. Plan on 32 GB for that, or 48 GB if you also want other apps open.

Is MiniCPM-V 4.7 on Ollama?

Yes, as openbmb/minicpm-v4.7, not in the main library.

How fast is MiniCPM-V 4.7 on an M4 Pro?

Collected oMLX runs show 75 to 82 tokens a second at 4-bit on a 48 GB M4 Pro, with about 21 GB peak memory.

Is the 1B model good enough?

For image description on small machines, maybe. In Micdrop’s agent test it never called a tool and invented answers instead, so I would not use it for agents.

Related local notes

Sources and what I did not test

I have not run MiniCPM-V 4.7. The speeds are other people’s runs, the quality claims are OpenBMB’s unless a tester is named, and the sizes were read on 4 November 2026.

Sources

Suhail Mohebi · published 4 November 2026 · the cover is a generated illustration