How should a M5 Mac Mini compare to a M4 mac mini for memory bandwidth and real world LLM usage?

[FlyingTexan]

I'm thinking of buying a few Mac Mini's to run LLMs. I have a M4 base model that's sitting in a drawer (company buy) and I'm running openclaw on an unraid server but I want to move to running my own LLMs vs using Claude Sonnet, Grok 4.1, etc. From what I've read you can mix/match the mac minis. No clue how it work but again I have a M4 base with 16GB sitting here. So I'm considering either buying 4 or so M4 models or going to wait for the M5. I have about $2k in apple credit and enough credit card points to get another $3k or so in minis before I have to pay anything out of pocket. So I'm open to ideas. I now total memory is the king thing to consider but not sure how much a M5 vs a M4 with memory bandwidth will affect things.

 

[FreakinEurekan]

You're making an assumption that there will BE a M5 mini... Apple never offered M3 in the mini, it's certainly conceivable that they'd skip M5 and wait for M6.

But, let's assume it is coming. Now we need to guess at memory bandwidth. Looking at the MacBook Pro, the M4 is 120GB/s and the M5 is 153GB/s. So pretty significant as far as that goes, a 27.5% improvement.

Or - go with M4 Pro, which is 273GB/s. That's literally a 127.5% improvement over M4, and available today!

 

[FlyingTexan]

You're making an assumption that there will BE a M5 mini... Apple never offered M3 in the mini, it's certainly conceivable that they'd skip M5 and wait for M6.

But, let's assume it is coming. Now we need to guess at memory bandwidth. Looking at the MacBook Pro, the M4 is 120GB/s and the M5 is 153GB/s. So pretty significant as far as that goes, a 27.5% improvement.

Or - go with M4 Pro, which is 273GB/s. That's literally a 127.5% improvement over M4, and available today!

Given the GPU increase in the M5 pro there's going to be something offered.

 

[OldMike]

I don't think a base M5 Mac Mini is going to provide a lot more useful performance than a base M4 Mac Mini. The M5 Pro and Max should have a greater improvement for local LLMs.

I have a few base M4 Mac Minis, a MBP M4 Pro, Mac Studio M4 Max and Mac Studio M3 Ultra.

In comparison to models that you are running, like Claude Sonnet, locally hosted models are quite inferior. Using these lower end models are even more problematic when running them in an agentic fashion. These are my opinions, from a developers perspective, who has been running LLMs locally (though beginning to use paid models more frequently).

I don't consider the models that can be run on the base Mac Mini to be very useful or performant for most serious work.

The minimum machine level for local LLMs, I feel is the M4 Pro. My MBP M4 Pro only has 24GB RAM, which is too little IMO.

I also feel the Mac Studio Mx Max is really the best deal going for running LLMs locally.

That being said, even my M3 Ultra with 96GB RAM doesn't have enough RAM to run open models that can begin to come close (though still a level below) the frontier models. I feel like a high quality, low cost model, like Gemini 3 Flash is a more cost effective way to go as opposed to investing in hardware solely for LLM usage.

Don't get me wrong, I'm not saying it is not fun or useful to run models locally. Local models continue to improve, and with MoE models they are becoming much faster to run locally as well.

I'm just saying, from my perspective, that you should temper your expectations and perhaps not try to chase hardware that might not get you where you want to be.

 

[leman]

I don't think a base M5 Mac Mini is going to provide a lot more useful performance than a base M4 Mac Mini. The M5 Pro and Max should have a greater improvement for local LLMs.

I have a few base M4 Mac Minis, a MBP M4 Pro, Mac Studio M4 Max and Mac Studio M3 Ultra.

In comparison to models that you are running, like Claude Sonnet, locally hosted models are quite inferior. Using these lower end models are even more problematic when running them in an agentic fashion. These are my opinions, from a developers perspective, who has been running LLMs locally (though beginning to use paid models more frequently).

I don't consider the models that can be run on the base Mac Mini to be very useful or performant for most serious work.

The minimum machine level for local LLMs, I feel is the M4 Pro. My MBP M4 Pro only has 24GB RAM, which is too little IMO.

I also feel the Mac Studio Mx Max is really the best deal going for running LLMs locally.

That being said, even my M3 Ultra with 96GB RAM doesn't have enough RAM to run open models that can begin to come close (though still a level below) the frontier models. I feel like a high quality, low cost model, like Gemini 3 Flash is a more cost effective way to go as opposed to investing in hardware solely for LLM usage.

Don't get me wrong, I'm not saying it is not fun or useful to run models locally. Local models continue to improve, and with MoE models they are becoming much faster to run locally as well.

I'm just saying, from my perspective, that you should temper your expectations and perhaps not try to chase hardware that might not get you where you want to be.

I second this. While M5 brings a lot of improvements specifically targeted at running ML models, at the end of the day the hardware is simply not large enough to be practically useful. Unless you are only interested in smaller models, in which case an improvement between 2x-4x is not impossible.

 

[FreakinEurekan]

it's certainly conceivable that they'd skip M5 and wait for M6

I don't get tired of being right! 🤣