A local language model has to fit in memory. If it doesn’t, nothing else on the spec sheet matters, not the chip and not the screen.
That’s why I think most coverage of the new OLED MacBook Pro is asking the wrong question. The 14-inch starts at $2,899 with an M5 Pro, 24GB of memory and 1TB of storage, and the extra $400 over the regular model buys a better panel and touch. For running models on your own machine, neither of those does anything. The memory tier does almost everything.
So this is the setup I’d build on it, step by step, and what I’d buy before building it. One thing I should say clearly: I haven’t done this on one of these laptops, because they don’t reach stores until November 6. What follows rests on model-size arithmetic and on benchmarks other people ran on the same chips, and I’ll name each source as I go. A few of those sources are enthusiast blogs, not labs, and they don’t always agree with each other.
The memory math comes first
The model itself is just a file full of numbers called weights. File size is the number of weights times the bytes each one takes. At full precision that’s two bytes per weight, so a 70-billion-parameter model would need around 140GB. Almost nobody runs it that way. Quantization squeezes each weight down to about half a byte (4-bit), and the same 70B model drops to roughly 40GB once you add overhead. That’s my own arithmetic, not a benchmark, but it’s close enough for planning.
Real files back it up. BestLLMfor’s M5 Max test lists Gemma 4 12B at 7.6GB and Qwen 3.8 27B at 17GB, both at 4-bit. OpenAI built its 120B open model, gpt-oss-120b, to run on a single 80GB GPU, so plan on somewhere above 60GB for anything in that class.
And not all of your RAM is available to the model. As far as I know macOS limits the GPU to roughly two-thirds or three-quarters of total memory by default, depending on the machine, and there’s a setting to raise it, but I wouldn’t push a laptop that’s also running a browser and an editor right to the edge. Long prompts add more on top, because the model keeps a running cache of everything it has read. So my rough tiers, and these are estimates, go like this. 24GB handles models up to about 14B comfortably and a 27B only just. 48GB runs 27B to 32B models with room for long prompts. 64GB can hold a 4-bit 70B, but it’s tight. And 128GB is the only tier where 120B-class models make sense.
The M5 Pro tops out at 64GB. For 128GB you need the M5 Max, and on the OLED model that starts at $4,499 before you touch the memory option.
Memory prices are what make this annoying. At the March launch, going from 24GB to 48GB on a 14-inch cost $400, then $200 more to reach 64GB and another $1,000 to reach 128GB, according to 36Kr. NewsBytes later reported Apple had raised these, with 48 to 64GB at $400 and the jump to 128GB at $2,000. I couldn’t find upgrade prices for the OLED models specifically, so check Apple’s configurator before trusting any of my numbers. Side note, and it’s a tangent: in one 9to5Mac piece, Apple’s price to go from 64GB to 128GB was $800 while a 64GB kit of desktop DDR5 cost about $700. Apple looked pricey then, and the shortage has made everyone else look pricey too.
The setup I’d build, in order
I’d start with Ollama because it’s the least work. You install it, run one command to pull a model, and you get a chat in the terminal. It also starts local server on your machine that other apps can talk to. BestLLMfor’s guide says Ollama now has an MLX engine, which means models run through Apple’s own machine learning library instead of a generic one, and that matters for speed a bit later.
For the first model I’d pick a mixture-of-experts one. These are models where only a few billion of the weights wake up for each word, so a 35B model can feel as quick as a 3B one. In BestLLMfor’s six-model test, a 35B model of this type, Qwen 3.6 35B-A3B, was the fastest by a long way.
Second, if speed starts to bother me, I’d try LM Studio, which has a friendlier model browser and can run MLX builds. Same models, different engine. I’d keep both installed, since it costs only disk space.
Third, I’d point my code editor’s assistant at the local server. Most of them accept a custom address, and it’s localhost plus a port number. This is where long prompts start to hurt, because a coding assistant sends your files again on every turn, and that brings us to prefill below.
And I’d keep Activity Monitor open for the first week. The memory pressure graph is the honest answer to “did I buy enough RAM?” Green is fine, yellow means the Mac is starting to swap, and red means the model is too big for the machine.
Speed is mostly bandwidth
Generation speed has a simple rule behind it. Each word the model writes requires reading the active weights from memory once, so the speed limit is memory bandwidth divided by the size of what gets read. A guide on InsiderLLM lists the M5 Pro at 307GB/s and the M5 Max at 614GB/s, and warns that some 14-inch M5 Max configurations are reported at only 460GB/s, so the GPU core count matters as much as the chip name.
Run the rule on the 27B model. It weighs 17GB, 614 divided by 17 is about 36, and BestLLMfor measured 36.6 tokens per second on it in the standard GGUF format. The ceiling works. A mixture-of-experts model only reads its few active billions per word, so the ceiling is much higher: the same test got 167 tokens per second on the 35B-A3B model, 113 on gpt-oss 20B and 79 on gpt-oss 120B. OpenClawDC reports 64 to 88 tokens per second on gpt-oss 120B and 55 to 66 on Qwen3.5–122B for the same chip, so two different sites at least land in the same neighbourhood.
The M5 Pro, with half the bandwidth, should be roughly half as fast for the same model. PromptQuorum says exactly that.
Now the part that took me longest. BestLLMfor reports Qwen 3.8 27B at 68 tokens per second in MLX against 36.6 in GGUF, a gap of 86 percent. Its own separate guide on the M5 says MLX is 30 to 40 percent faster. Both can’t be the normal case. And my bandwidth math says a plain 17GB model on 614GB/s shouldn’t pass about 36 tokens per second whatever engine reads it, so a 68 means either the MLX file is smaller or something like speculative decoding was switched on, and the post doesn’t say which. I spent way too long trying to make those three numbers agree and I couldn’t. So I’d plan around 30 to 40 percent until someone explains the 68. The test, in case you want to check it, was run on September 16 with Ollama 0.34.1 on a 128GB M5 Max with the 40-core GPU.
Prefill is the number nobody publishes, and it’s the one you feel.
Prefill is the time the model spends reading your prompt before it writes anything. It’s limited by compute, not bandwidth, so my bandwidth rule is only half the story. That’s mostly true of generation, though I’m oversimplifying a bit, since longer contexts slow generation too. BestLLMfor measured 1,388 tokens per second of prefill on a 16,689-token document, about 12 seconds of reading before 38 seconds of answer. A guide on InsiderLLM lists Apple’s claim as up to four times faster prompt processing than M4, and the M5 GPUs now carry a Neural Accelerator in each core, which looks like the reason. Do the arithmetic on a coding agent that re-reads 30,000 tokens each turn and you get about 22 seconds of waiting before it types a single character. That’s livable. At half the prefill speed it isn’t.
Then there’s heat, which nobody has tested on the new body at all. Notebookcheck’s March review of the 14-inch M5 Max reported serious throttling and CPU results that jumped around, and said the 16-inch held up better. OpenClawDC notes the fans are audible under long loads. PromptQuorum puts power draw during generation at 60 to 100 watts for the M5 Max and 25 to 45 for the M5 Pro. At those numbers Apple’s 24-hour battery claim stops being relevant, so heavy local AI is a plugged-in job. Apple says the new hinge cooling keeps performance the same as before. If “the same” includes the 14-inch throttling, a long agent run is exactly where it would show.
What I’d actually buy
Two other laptops come up in every local-AI thread, so they need a mention. The first is the ASUS ROG Flow Z13 with AMD’s Ryzen AI Max+ 395. OpenClawDC lists the 128GB version at around $2,800 in August 2026, with about 256GB/s of bandwidth, in a 13-inch tablet body with modest sustained power. That’s the same 128GB for far less money. It’s also less than half the Mac’s bandwidth, so expect generation to be correspondingly slower on the same model. The second is a gaming laptop with an RTX 5090. As far as I understand, those carry 24GB of video memory, which holds a 27B model at 4-bit with room to spare and can’t hold a 40GB 70B one, so the rest spills into slower system memory. It’s fast on anything that fits. The ceiling is the problem.
On the Apple side, the M6 14-inch at $1,999 starts with 16GB, which is fine for 7B to 12B models as a toy and not for a serious setup. The regular non-touch M5 Pro 14-inch at $2,499 comes with 24GB, and at March upgrade prices, 48GB would add $400. That’s about $2,899, the same money as the OLED base with 24GB, give or take whatever Apple has changed since.
So here’s my lean, around 90–10. If you want local AI for coding help and document chat, skip the OLED and buy the 48GB non-touch M5 Pro. Yes, the OLED screen is nicer to look at all day. That’s true and it isn’t what this article is about. If you know you’ll run 70B-class models or bigger, you need the M5 Max with 128GB, and I’d take the 16-inch because of the throttling reports. The memory upgrade alone is somewhere between $1,000 and $2,000 depending on which report you believe, and at that point it’s worth asking whether a Strix Halo box would do.
Before paying for any extra memory, I’d do one thing. Install Ollama on whatever Mac or PC you have now, pull a 30B-class mixture-of-experts model that fits, and use it for a real week of work. If you stop opening it after three days, 48GB is more than enough. If you keep hitting the size limit, you’ll know exactly which tier to pay for, and the memory pressure graph will tell you.