YouTuber runs DeepSeek‑V4‑Flash on MacBook M5 Max, proving heavy LLMs can run on Apple silicon
- Tech‑Practice posted a new YouTube benchmark showing DeepSeek‑V4‑Flash running fully local on a MacBook M5 Max, using custom llama.cpp instead of cloud APIs. (youtube.com) - The telling detail is size: the creator says the model footprint is about 85 GB, while the llama.cpp fork targets 128 GB Macs with 2‑bit expert quantization. (youtube.com) - It matters because DeepSeek‑V4‑Flash just landed as open weights, and early tooling is already pushing “laptop‑class” local inference further than expected. (huggingface.co)
A laptop benchmark is not the same thing as a product launch. But this one matters anyway. A YouTuber called Tech‑Practice just showed DeepSeek‑V4‑Flash running fully lo(youtube.com) in the loop. That is the interesting part — not because one video proves anything universal, but because it turns a fuzzy claim about “Apple silicon can run big models” into a concrete test. (youtube.com) ### What actually ran? The video says DeepSeek‑V4‑Flash ran on an Apple M5 Max MacBook through a custom llama.cpp build, and th(huggingface.co)not a toy 7B or 14B model. DeepSeek‑V4‑Flash is part of DeepSeek’s freshly released V4 family, and the public model listings put Flash at 284B total parameters in one variant, with open weights now live on Hugging Face. (youtube.com) ### Why is the Mac part the story? Apple’s pitch for local AI has always been unified memory. Basically, the CPU, GPU, and neural hardware can all wor(youtube.com) across separate VRAM and system RAM islands. That does not make a MacBook magically faster than a rack of GPUs. But it does make “can this even fit?” a much more interesting question on high‑memory Macs than on most laptops. The experimental DeepSeek V4 Flash llama.cpp fork says the target is Macs with 128 GB of RAM, using 2‑bit quantization on routed experts. (github.com)ause raw model size is the brick wall. DeepSeek‑V4‑Flash is huge in full form, so nobody is casually dropping the untouched weights onto a notebook and calling it a day. The trick is aggressive compression — especially quantizing the expert weights hard enough that the model becomes loadable without collapsing completely. Think of it like folding a giant paper map until it fits in your jacket pocket. You lose some neatness, but now you can actually carry it. The 85 GB figure in the video sits right in that practical zone — still heavy, but no longer absurd for a premium Mac. (youtube.com) ### Does this mean any Mac can do it? No. That is the catch. This is a MacBook M5 Max test, and the repo behind the experiment explicitly aims at 128 GB Macs. So the headline is not “Apple laptops now run frontier models.” The headline is narrower: one of the heavier new open models can be coerced into usable local inference on very expensive Apple silicon if the software stack is clever enough. That is still a meaningful shift. (youtube.com) ### Why DeepSeek‑V4‑Flash specifically? Because Flash is the practical member of the new V4 family. DeepSeek’s V4 collection includes bo(youtube.com)ying to bend into local workflows. The model card and collection pages show DeepSeek pushing V4 as open weights with long context, while community ports and quantized variants appeared almost immediately. That fast tooling response is part of the story — the model did not just launch, it started getting adapted for real machines right away. (huggingface.co) ### So what change(youtube.com)is capable. People have been doing local inference on Macs for a while. The new thing is that the ceiling moved. DeepSeek‑V4‑Flash only showed up days ago, and already there is an experimental llama.cpp fork, community quantization work, and a public laptop demo of a roughly 85 GB local run. That compresses the gap between “released model” and “someone actually using it on a desk” to almost nothing. (github.com) ### Bottom line? This does not prove a MacBook is the new default (huggingface.co)ting factor is less “is local inference possible?” and more “how much memory, how much quantization, and how much patience do you have?” For Apple, that is a pretty good place to be. (youtube.com)
