MacBook Pro and local LLM

[al404]

I own a MacBook Air M3 16Gb 15" and really like it except for local LLM
I mostly use Claude Code but since AI market is unpredictable I was looking into LM studio

Air is not the right machine to run local LLM it get gets hot quite fast

I was wondering if a MacBook PRO M5 32Gb could be good enough to run some decent model
I actually prefer the 15" air size and I guess I would have to downgrade to 14"

Not sure if is I should wait and see how Cloud AI will evolve or upgrade know to something that let me have a local option

 

[Bigwaff]

Local LLM for what use cases?

 

[al404]

Text and coding I'm just starting to mess a round with them on my Mac mini M4 24Gb and I see I can run Gemma 4 26B, Qwen3.5-27B-Claude-4.6-Opus ( but is slow), GPT oss 20b

Is pretty hard to understand how good they are but from what I'm reading Gemma 4 seems pretty smart for his RAM occupation

 

[splifingate]

Qwen3.5-27B

Using this model can bump one of my GPUs with 32GB of dedicated VRAM in linux to a full 300W...that's going to be a challenge for any laptop.

 

[benholt]

32GB on an M5 Pro is just barely enough to be useful for 20–30B coding models, but you'll feel the RAM ceiling fast once context climbs past ~16k tokens. The model weights eat most of your headroom and macOS starts getting unhappy when wired memory creeps past 80%. On my 16" M3 Max (64GB) Qwen3 30B and Gemma 4 27B at q4 are comfortable with decent context — bumping the Qwen to a 32k window is already brushing into swap territory, so your 24GB mini is going to feel really tight once anything 20B+ is loaded. The 14" M5 Pro will also throttle harder than your 15" Air under sustained generation, since these models keep both the CPU and GPU pegged for a while. If local LLM is the actual reason for the upgrade, I'd stretch to 48GB or wait for an M5 Max — that's the real comfort floor for the model sizes you're naming, and it'll age a lot better than a 32GB Pro will.

 

[al404]

The 14" M5 Pro will also throttle harder than your 15" Air under sustained generation, since these models keep both the CPU and GPU pegged for a while. If local LLM is the actual reason for the upgrade, I'd stretch to 48GB or wait for an M5 Max — that's the real comfort floor for the model sizes you're naming, and it'll age a lot better than a 32GB Pro will.

Thanks I got to the conclusion that I should get a M5 Pro with Pro chip not sure if 14 or 16 and with 64Gb or 48Gb. Of course 64 would be the best configuration.
Why are you saying that "The 14" M5 Pro will also throttle harder than your 15" Air" I thought Air will always throttle more since is fanless.

An other option can be to keep my MacBook M3 16gb and just use open models in the cloud

I guess I will start testing then in cloud and see what I could really get with Qwen3.6 35b 6bit

 

[carl varley]

I would look at a used 64gb m1 or m2 max. I got the 14 inch with 64gb with just 24 cores for less than £900 and it can run some of the larger models quite well. All faster than i can read. I mostly use Jan, but just started to use omlx and it's a real tps boost. If cost is not a problem then just get the m5 max with as much ram as you can afford.

 

[keksikuningas]

I run 27B-8bit on my 14inch 96GB M3 Max. Speed is decent but laptop really heats up. I recommend going with at least 64gb to support a decent context + all the tools you'd be running alongside. There should be benchmarks somewhere for M5 Pro vs M5 Max. Sadly Apple bumped up the price for all Macs, but maybe your local retailer hasn't updated the prices yet