Does MiniCPM-V 4.7 really run in 3GB of RAM?

OpenBMB reports a 3GB working-memory footprint for MiniCPM-V 4.7 on an iPhone 17 Pro Max, but the document does not define whether that is peak or average or what components it covers. The 4-bit download is 21.2GB, so most weights remain in flash storage rather than RAM.

How fast is MiniCPM-V 4.7 on an iPhone?

OpenBMB reports 13+ tokens per second on an iPhone 17 Pro Max. The only published stage-by-stage table was measured on a 16GB iPad Pro with the M4 chip and runs from 2.97 to 17.39 tokens per second. The iPhone figure has not been independently reproduced.

What is Lookahead Expert Scheduling in MiniCPM-V 4.7?

LES is OpenBMB's prefetch scheme that uses the previous layer's router signal to predict the next layer's experts and load them from flash while the current layer computes. Missed experts load through the normal path, so the guess does not change the model's routing.

How much storage does MiniCPM-V 4.7 need?

OpenBMB's GGUF repository lists the 4-bit Q4KM file at 21.2GB, Q8 at 36.9GB and F16 at 69.4GB. Plan disk space around the file you intend to download, not the memory figure in the headline.

Has anyone outside OpenBMB measured MiniCPM-V 4.7 on a phone?

No independent reproduction of the iPhone throughput was published as of October 2026. Related research on a different system measured a confidence-gated prefetch predictor at 61.31% accuracy with minimal speedup on an iPhone, which shows prefetch accuracy is the variable that decides whether this approach pays off.

MiniCPM-V 4.7 on iPhone: What 3GB Really Means

MiniCPM-V 4.7 on iPhone is best understood as a flash-resident model with a small working set in RAM, not a 35B model compressed into 3GB. OpenBMB's own GGUF repository lists the 4-bit Q4_K_M download at 21.2GB, while the iPhone figure covers only what the device holds in memory at once.

What does the MiniCPM-V 4.7 on iPhone 3GB claim actually measure?

The 3GB in OpenBMB's MiniCPM-V 4.7 on iPhone claim is a reported working-memory footprint, not the size of the model. The 35.2-billion-parameter checkpoint at 2 bytes per parameter is 70.4GB, and the 4-bit Q4_K_M file in OpenBMB's GGUF repository is 21.2GB, so most weights must live somewhere other than RAM.

OpenBMB's model card, README and LES document do not say whether 3GB is peak or average memory. They do not state whether it includes the vision encoder, the KV cache of earlier tokens, or the draft model used for speculative decoding. Treat the number as a headline with an unstated scope.

Two numbers ending in three cause most of the confusion. The model activates roughly 3 billion parameters per token, and OpenBMB reports a 3GB memory footprint. Those measure different things: compute per token versus resident memory.

How should you plan storage and evaluate the quality numbers?

Plan storage around the nearest published file size, roughly 21.2GB for the 4-bit Q4_K_M download in OpenBMB's GGUF repository, rather than around 3GB. The memory figure does not describe disk use.

OpenBMB's launch post reports 90.4 on OCRBench at 16x visual token compression, which it places 5.2 points ahead of a 35B Qwen model. Treat that as a vendor number until independent evaluation appears.

Frequently asked questions about MiniCPM-V 4.7 on iPhone

Does MiniCPM-V 4.7 really run in 3GB of RAM? OpenBMB reports a 3GB working-memory footprint for MiniCPM-V 4.7 on an iPhone 17 Pro Max, but the document does not define whether that is peak or average or what components it covers. The 4-bit download is 21.2GB, so most weights remain in flash storage rather than RAM.

How fast is MiniCPM-V 4.7 on an iPhone? OpenBMB reports 13+ tokens per second on an iPhone 17 Pro Max. The only published stage-by-stage table was measured on a 16GB iPad Pro with the M4 chip and runs from 2.97 to 17.39 tokens per second. The iPhone figure has not been independently reproduced.

What is Lookahead Expert Scheduling in MiniCPM-V 4.7? LES is OpenBMB's prefetch scheme that uses the previous layer's router signal to predict the next layer's experts and load them from flash while the current layer computes. Missed experts load through the normal path, so the guess does not change the model's routing.

How much storage does MiniCPM-V 4.7 need? OpenBMB's GGUF repository lists the 4-bit Q4_K_M file at 21.2GB, Q8 at 36.9GB and F16 at 69.4GB. Plan disk space around the file you intend to download, not the memory figure in the headline.

Has anyone outside OpenBMB measured MiniCPM-V 4.7 on a phone? No independent reproduction of the iPhone throughput was published as of October 2026. Related research on a different system measured a confidence-gated prefetch predictor at 61.31% accuracy with minimal speedup on an iPhone, which shows prefetch accuracy is the variable that decides whether this approach pays off.

Turning a dense video breakdown into a readable article

A video like this one packs a file-size table, a stage-by-stage benchmark, a paper comparison and a set of open questions into eleven minutes. Written out, the same reasoning becomes something a reader can check line by line, with sources attached to each number.

If you already explain technical work on YouTube, that explanation can become a written article. Skala blog takes a YouTube URL, transcribes the video and generates a structured draft you review before publishing. Teams that want the same workflow wrapped in a ready-made TypeScript stack can look at Crazystack typescript, built by Dev doido and documented at crazystack.com.br.

The MiniCPM-V 4.7 story rewards that kind of care. The mechanism is credible, the iPhone number is still vendor-reported, and the difference between 3GB and 21.2GB decides how you plan storage.

Skala Blog

Source video