don't run on m1 MacBook iPad simulator and iPhone 13 real device

[Lxdro]

I don't speak Chinese so maybe I did something wrong and didn't really understood the app but I successfully downloaded it and ran it first on my iOS 18 iPhone 13 (physical device), downloaded the 4 bit quantized model, and try to send a prompt. It loaded endlessly pending for models answer. Then I tried on my MacBook m1 16gb ram on any iPad simulator, followed the same process, and I also couldn't manage to load any output from the model.
Do you know minimum spec to have maybe?

[avolstudio]

I've got the same issue withj Iphone 14Pro

Downloaded the app, downloaded models in settings view, but nothing is generated.
Did you solve this issue? @Lxdro

[Lxdro]

I've got the same issue withj Iphone 14Pro

Downloaded the app, downloaded models in settings view, but nothing is generated. Did you solve this issue? @Lxdro

no sadly, I just think our devices don't have enough computation power

[tc-mb]

Thanks for the report — to triage this properly, could you tell us:

1. Which model version did you download? The app currently offers three:

   • MiniCPM-V 2.6 (8B, ~5.4 GB, needs ≥ 8 GB RAM device)
   • MiniCPM-V 4.0 (4.1B, ~2.9 GB, ≥ 6 GB)
   • MiniCPM-V 4.6 (1.3B, ~1.6 GB, ≥ 6 GB)

   This matters a lot — iPhone 13 / 14 Pro are 6 GB devices, V2.6 won't fit on them but V4.0 / V4.6 should be fine.

2. App version (TestFlight build number, shown in Settings).

   The latest build adds a ModelScope mirror as fallback when HuggingFace isn't reachable, plus several stability fixes — please test on: https://testflight.apple.com/join/yNKyFZwW

3. Any logs from Xcode console / Console.app when you tap send?

   Specifically the lines starting with llama_model_load_from_file_impl, ggml_metal_init, or mtmd_. "Loading endlessly" without any error line usually means either:

   • the model context failed silently (RAM pressure), or
   • the prompt is still being prefilled on a large image — V2.6 prefill on a 6 GB device can take 30s+ before the first token.

Side note on simulator: this demo is only validated on physical iOS devices. iPad simulator on M1 Mac uses a different mmap / Metal path and frequently fails to load GGUF or compile the Metal shader library — please test on a real device.

If you can share the model version + a few lines of the console log, we can pinpoint whether it's RAM, network, or a model-side issue.