Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
May I run DeepSeek on Mac mini 3,1?
- Thread starter mikezang
- Start date
-
- Tags
- deepseek macmini 2009
- Sort by reaction score
There are various sizes of Deepseek R1 on Ollama - from 1.5B to 671B parameters.
Whether you can run them on a late 2009 Mac mini is probably something no one else has tried. I’ve tried running small LLMs on my late-2014 Mac mini and it did work, albeit slowly (I’ve not tried Deepseek yet).
Whether you can run them on a late 2009 Mac mini is probably something no one else has tried. I’ve tried running small LLMs on my late-2014 Mac mini and it did work, albeit slowly (I’ve not tried Deepseek yet).
I remember trying to run a large Lisp application on a Vax 730 in the mid-80. I was the sole user of the machine.
I had a portable teletype with a clackity keyboard and printout for feedback logged in via 300 Baud modem,
I am not making this up!
I ran it from my pool while floating on an inflated mattress. Every five minutes, I would hear the printer responding with some clacky sounds. I'd float back, enter my next line, hit return, and float some more.
Restful, but not productive. I imagine a late mini 2009 will do similar, trying to run a model on a tiny Intel chip that is 15 years old.
BTW, Ollama requires Big Sur.
I had a portable teletype with a clackity keyboard and printout for feedback logged in via 300 Baud modem,
I am not making this up!
I ran it from my pool while floating on an inflated mattress. Every five minutes, I would hear the printer responding with some clacky sounds. I'd float back, enter my next line, hit return, and float some more.
Restful, but not productive. I imagine a late mini 2009 will do similar, trying to run a model on a tiny Intel chip that is 15 years old.
BTW, Ollama requires Big Sur.
Most likely you cannot, at least officially.
Mac mini late 2009 is only supported by macOS 10.13 (El Captan). But Ollama requires 11.0 (Big Sur) or above. LM studio only supports Apple Silicon macs.
If you really need to, you can compile llama.cpp by yourself and do all the chores of model conversion, etc.
But the CPU is a bottleneck. The smallest model for DeepSeek r1 is 1.5B parameters. The Intel Core2 Duo might be too slow for that. It doesn't even support SSE. Without vector instructions, the perf could be disastrous.
Mac mini late 2009 is only supported by macOS 10.13 (El Captan). But Ollama requires 11.0 (Big Sur) or above. LM studio only supports Apple Silicon macs.
If you really need to, you can compile llama.cpp by yourself and do all the chores of model conversion, etc.
But the CPU is a bottleneck. The smallest model for DeepSeek r1 is 1.5B parameters. The Intel Core2 Duo might be too slow for that. It doesn't even support SSE. Without vector instructions, the perf could be disastrous.
Note the Intel Core2 Duo supported up through SSE4.1. However, it didn't support AVX/AVX2 (not to mention AVX-512) which might be the more common target for any assembly/instrinsic optimized code these days.
I see that wasn't listed on the Intel spec page you linked to but I am guessing that page is incomplete for old processors. SSE4.1 came out circa 2007 and prior SSE/SSE2/etc instructions were standardized years before that.
Intel Core 2 - Wikipedia
But agree with your summary. I think likely possible as a proof of concept/test of strength exercise (perhaps not far behind running Linux on an 8088) possibly requiring rebuilding an entire software stack on an older macOS and/or hacking an older macOS onto that hardware. Or perhaps switching it to Linux. But at that point, a Beelink box likely a better result for this in much less time for not much money.