Apple M5 for Local AI: MacBook Pro M5, M5 Pro & M5 Max (128GB) Guide

Short answer: if you want a laptop for serious local AI, buy the MacBook Pro M5 Max with the 40-core GPU and 128GB of unified memory (614GB/s; from $4,699 for the 14-inch or $4,999 for the 16-inch before the 128GB upgrade, US Apple Store, November 2, 2026). Only the 40-core M5 Max takes 128GB; the 32-core version is fixed at 36GB. It is no longer the only 128GB Mac: the Mac Studio M5 Max reaches 128GB on a desk for less, and the M5 Ultra Mac Studio goes to 512GB. A published review measured a 4-bit Llama 3.3 70B at 13.1 tok/s and the Qwen3 30B-A3B mixture-of-experts model at 131 tok/s on this chip. The base 14-inch is now the M6 (up to 170GB/s, 32GB max), an 8B-to-14B machine.

That is the whole recommendation. The rest of this guide is the evidence: what each tier actually costs and runs, published tokens-per-second numbers with sources, the setup commands, and the honest list of things a $4,700+ laptop still does worse than a $2,000 GPU. If you are choosing between Mac generations or between Mac and PC entirely, the Apple Silicon buying guide and Mac Studio vs PC comparison cover those calls.

Updated November 2, 2026: Apple changed the lineup after this guide was first published. The Mac Studio moved to the M5 Max and M5 Ultra (up to 512GB) on September 22, the base 14-inch MacBook Pro moved to the M6 on October 27, and on October 28 Apple added a MacBook Pro with PROLED XDR Display (M5 Pro from $2,899, M5 Max from $4,499, shipping November 6) alongside the standard models, which now cost more than at the March launch. The speed table now uses published measurements (the llama.cpp Apple Silicon thread and tbreak's M5 Max review) plus labelled bandwidth ceilings. The leaderboard we quoted in August now marks those rows as estimates, so they have been removed.

---

The M5 Lineup, in Local-AI Terms

The spec that matters is memory bandwidth, because token generation is bandwidth-bound, and the MacBook Pro tiers are 153-170, 307, and 460/614GB/s. Everything else on the spec sheet is secondary for LLM work.

Specs below are from Apple's MacBook Pro tech specs page and prices from the US Apple Store, both read November 2, 2026. Prices are for the standard-display models; the M5 Pro and M5 Max launched March 3, 2026, and the M6 replaced the base M5 on October 27. The MacBook Pro with PROLED XDR Display added on October 28 uses the same M5 Pro and M5 Max chips and costs more:

Three things worth knowing beyond the table:

• Every GPU core carries a Neural Accelerator in this generation. Apple's claim is "up to 4x faster LLM prompt processing than M4 Pro and M4 Max." That is a prefill claim, not a generation-speed claim; the benchmark section shows the difference.
• Base storage is large: M6 and M5 Pro models start at 1TB, M5 Max at 2TB. Model files are huge (a Q4 70B is ~43GB and you will hoard several), so this matters more than it sounds.
• You do not need the top storage tiers. 128GB of memory plus the stock 2TB is the sane AI build. Apple has moved memory pricing and availability more than once this year, so check the configurator for the current cost of the 128GB upgrade rather than trusting any number published even a month ago.

If you are deciding how much memory a specific model needs before you spec anything, our Apple Silicon AI calculator does the arithmetic for you.

---

Is the M5 Max Still the Only 128GB Mac?

No. For most of 2026 the M5 Max MacBook Pro was the highest-memory Mac Apple sold; since September 22, 2026 the Mac Studio with M5 Max also reaches 128GB and the M5 Ultra Mac Studio reaches 512GB. The laptop is still the only 128GB Mac you can carry.

It happened in three steps:

1. The big-memory Mac Studio tiers disappeared during the DRAM shortage. MacRumors reported that Apple pulled the M3 Ultra's 512GB tier in March 2026 and the 256GB and 128GB tiers by May, leaving the M3 Ultra fixed at 96GB. It was the same memory squeeze that has pushed GPU street prices 50%+ over MSRP. Through the summer the M5 Max MacBook Pro was the only 128GB Mac.
2. Apple announced the Mac Studio with M5 Max and M5 Ultra on August 25, 2026, shipping September 22. The Mac Studio M5 Max starts at $2,499 (32-core GPU, 36GB); the 40-core version starts at $3,099 with 64GB and configures to 128GB. The M5 Ultra starts at $5,499 with 96GB at 1.2TB/s; 256GB and 512GB need the 80-core version, from $6,799 (US Apple Store, November 2, 2026).
3. Apple added a MacBook Pro with PROLED XDR Display on October 28, alongside the standard models and shipping November 6. It uses the same M5 Pro and M5 Max chips with the same 128GB ceiling, from $2,899 (14-inch M5 Pro) to $4,799 (16-inch M5 Max) per Apple's announcement. The standard models also cost more than at the March launch: the 14-inch M5 Max now starts at $4,099, against $3,599 in March.

So the question is no longer whether Apple sells 128GB, but where you want it. On a desk, the Mac Studio M5 Max gets the same 40-core, 614GB/s chip with 128GB for less money than the laptop, and the M5 Ultra doubles the bandwidth. The MacBook Pro M5 Max earns its premium only if the machine has to travel. (The M3 Ultra Mac Studio is no longer sold new; its 819GB/s still beats the M5 Max's 614GB/s on models that fit in 96GB, which makes a refurbished one worth a look.)

---

Real Benchmarks: What an M5 Max Actually Generates

Published runs on a 40-core, 128GB M5 Max put a 4-bit Llama 3.3 70B at 13.1 tok/s, the Qwen3 30B-A3B mixture-of-experts model at 131 tok/s, and the llama.cpp 7B Q4_0 reference test at 119.9 tok/s. Two named sources, both dated: tbreak's September 21, 2026 review of a Mac Studio M5 Max (the same 40-core, 614GB/s chip as the top MacBook Pro) and the community llama.cpp Apple Silicon thread on GitHub. Where nobody has published a run, the table gives a labelled bandwidth ceiling (614GB/s divided by the Ollama file size) instead of a guess.

Sources: tbreak, Mac Studio M5 Max review (September 21, 2026): Mac Studio with 18-core CPU, 40-core GPU and 128GB, MLX 0.32.2 / mlx-lm 0.31.3 and Ollama. llama.cpp Apple Silicon performance thread (GitHub discussion #4167, read November 2, 2026): text generation (tg128) on LLaMA 7B Q4_0. Ceilings are 614GB/s divided by the file size Ollama lists; real runs land below them. The 235B row stays because readers ask about it: qwen3:235b is 142GB at Q4, more than 128GB of RAM, so it needs a ~3-bit quant, and no M5 Max run of it has been published.

How to read that table honestly:

• MoE models are the sweet spot. Qwen3 30B-A3B runs at 131 tok/s because only ~3B parameters activate per token, and GPT-OSS 120B (65GB on disk) still generates 75.7 tok/s, several times a dense 70B's 13.1. On a 128GB machine, MoE models are what the capacity is for.
• Dense 70B at ~13 tok/s is reading speed, not build speed. It is also within about a token of the ~14 tok/s ceiling, so no runtime will make it much faster on this chip. Fine for chat, agents, and code review; painful for batch jobs.

Versus the M4 Max: in the llama.cpp thread the 40-core M5 Max generates 119.9 tok/s on the 7B Q4_0 test against the 40-core M4 Max's 83.1, and processes the prompt at 3,220 tok/s against 886, about 3.6x. Two cautions: the M5 rows were run on a newer llama.cpp build than the M4 rows, so part of the gap is software; and the bandwidth bump from 546 to 614GB/s (+12%) caps what the chip itself adds to generation on large models. The prefill jump matches Apple's claim of up to 4x faster LLM prompt processing than M4 Max, credited to the Neural Accelerator in each GPU core. In practice that means long-context work (pasting a codebase, RAG over big documents) is where M5 feels like a new machine. If you already own an M4 Max with 128GB and mostly chat with short prompts, stay put; this is not an upgrade generation for you.

---

Which Config to Buy

Buy the memory tier your target models need, then stop: 32GB for 8B-14B models, 64GB (M5 Pro or Max) for 32B-class, 128GB (40-core M5 Max only) for 70B+ and big MoE. Overspending on CPU cores or storage does nothing for tokens per second.

16GB / 24GB — the accidental AI Mac. If you have (or buy) a base M5 or the new M6 for other reasons, it runs 4B-8B models genuinely well: the ceiling for an 8B Q4 file (4.9GB) is ~31 tok/s at 153GB/s and ~35 tok/s at the M6's 170GB/s, and the llama.cpp thread lists the 10-core M5 at 31.9 tok/s on its smaller 7B Q4_0 test. That is still faster than most people read code. Our best LLMs for a 16GB Mac page ranks exactly what fits in that footprint.

32GB — the honest budget pick (14-inch M6 from $1,999 at 16GB, configurable to 32GB). A Q4 14B fits with room for context; a Q4 32B (20GB) technically loads, but its ceiling at 170GB/s is ~8 tok/s. The model picks for this tier are on our best LLMs for a 32GB Mac page.

48GB / 64GB M5 Pro ($2,499+ for the 14-inch). 307GB/s doubles the base chip's bandwidth: the llama.cpp thread lists the M5 Pro at 66 to 68 tok/s on the 7B Q4_0 test, and the ceiling for a 20GB Q4 32B is ~15 tok/s. 64GB needs the 20-core GPU version. A Q4 70B (43GB) squeezes into 64GB at a ~7 tok/s ceiling, so budget your patience. If the machine never leaves your desk, the Mac mini with the same M5 Pro starts at $1,699 (US Apple Store, November 2, 2026).

128GB M5 Max — the reason this page exists. Take the 40-core GPU: it is the only M5 Max that takes 128GB (the 32-core version is fixed at 36GB and 460GB/s). Take 128GB, keep the stock 2TB. The 14-inch starts at $4,699 and the 16-inch at $4,999 before the memory upgrade; the 16-inch has more thermal headroom for sustained inference, which is what LLM serving is. If it will live on a desk, price the Mac Studio M5 Max with the same 40-core chip first (from $3,099 at 64GB).

One warning from this year's pattern: Apple removed memory tiers from the Mac mini and Mac Studio mid-cycle during the DRAM crunch, then brought 512GB back with the M5 Ultra. Memory prices in the configurator have moved more than once in 2026, so check the current 128GB upgrade price before you budget.

---

Setup: Ollama and MLX on an M5

Two runtimes cover everything: Ollama for convenience and model breadth, MLX for Apple-optimized speed. Both are free; install both.

Ollama (easiest start — models in one command):

# Install (or download the app from ollama.com/download/mac)
brew install ollama

# Pick by memory tier
ollama run llama3.1:8b     # 16GB+ — fast on any M5 or M6
ollama run qwen3.6:27b     # 32GB+ — best small dense model
ollama run gemma3:27b      # 32GB+ — multimodal, 128K context
ollama run llama3.3:70b    # 64GB works, 128GB comfortable

MLX (Apple's own framework; tbreak ran its 70B and 30B tests on it):

pip install mlx-lm

# The mlx-community org on Hugging Face maintains Mac-ready conversions
mlx_lm.generate --model mlx-community/Llama-3.3-70B-Instruct-4bit \
  --prompt "Explain unified memory in one paragraph"

# Interactive chat
mlx_lm.chat --model mlx-community/Llama-3.3-70B-Instruct-4bit

Most of the published M5 Max figures above were run under MLX, and Apple credits the Neural Accelerators with speeding up its prefill.

Two practical tips for a 128GB machine:

• macOS caps GPU-wired memory below your total RAM (roughly 75% by default). The community-documented override is sudo sysctl iogpu.wired_limit_mb=110000 to let a 128GB Mac dedicate ~110GB to models. It resets on reboot; use it when a big MoE almost fits.
• Watch memory pressure, not free GB. Activity Monitor's memory pressure graph should stay green while a model is loaded. Yellow means swapping, and swapping murders tokens per second.

Full first-hour walkthrough — models, context settings, the works — in our Llama on Mac setup guide.

---

Honest Limitations

An RTX 5090 has about 2.9x the M5 Max's memory bandwidth (1,792GB/s vs 614GB/s) and generates proportionally faster on anything that fits in its VRAM, CUDA-only tooling still does not run, and Apple's "4x faster" AI claim is about prefill, not generation. Know what you are not buying:

• Generation speed is a bandwidth story Apple only partly fixed. 614GB/s is excellent for a laptop and roughly a third of a flagship GPU's GDDR7. Dense 70B at ~13 tok/s measured (~14 tok/s ceiling) is where the M5 Max stops; more bandwidth on a Mac now means the 1.2TB/s M5 Ultra on a desk. If your models fit in 24-32GB of VRAM, a discrete GPU is simply the faster tool.
• The 4x prompt-processing claim is real but narrow. Prefill got dramatically faster thanks to the per-core Neural Accelerators (3.6x on the llama.cpp thread's 7B test); generation on big models is capped by the 12% bandwidth bump over M4 Max. Reviews that only quote time-to-first-token flatter the machine.
• No CUDA. Fine-tuning tooling on Mac (MLX-LM does QLoRA) is workable for 7B-14B but years behind the CUDA ecosystem. Training seriously? Rent NVIDIA.
• The AI build is expensive and the price direction is up. $4,699 is the doorway for the 40-core chip, not the price: the 128GB config lands well north of that, and the 14-inch M5 Max starts $500 higher than at its March launch ($4,099 now, $3,599 then).
• Thermals are real on the 14-inch. Sustained inference is a space heater workload; the 16-inch chassis holds clocks better. Neither throttles catastrophically, but do not expect published numbers on battery in a quiet room — plugged-in, high-power mode is where those runs happen.

---

M5 Max vs the Alternatives

Buy the M5 Max MacBook Pro for portable capacity; buy a Mac Studio (M5 Max or M5 Ultra) for the same or more on a desk; buy a GPU tower for raw speed under 32GB. The one-table version:

The interesting fight is now laptop against desk with the same chip: a Mac Studio M5 Max with the 40-core GPU and 128GB has the identical 614GB/s for less money than the 14-inch MacBook Pro, so the laptop premium buys portability and a screen. Past 128GB the M5 Ultra is the only Apple option: at 256GB it holds qwen3:235b at Q4 (142GB), which no MacBook Pro can, and in the llama.cpp thread its 7B generation figure is 179.1 tok/s against the M5 Max's 119.9. The full desk-vs-tower argument is in Mac Studio vs PC for AI.

Against the PC, the math has not changed since the M4 generation: NVIDIA wins tokens per second, Apple wins model size per watt and per decibel. What changed in 2026 is the context — with GPU prices 50%+ over MSRP, the premium for going Mac shrank even as Apple raised its own prices.

---

Verdict

The M5 Max 128GB is the best local-AI laptop Apple has shipped; on a desk, the Mac Studio now does the same job for less, and the M5 Ultra goes further. The decisions, compressed:

1. Serious about big local models and need a laptop? 16-inch M5 Max, 40-core GPU, 128GB, stock 2TB. From $4,999 plus the memory upgrade.
2. Same models, desk only? Mac Studio M5 Max with the 40-core GPU and 128GB (from $3,099 at 64GB plus the upgrade), or the M5 Ultra for 1.2TB/s (from $5,499) or 256GB+ (80-core, from $6,799).
3. Daily driver is a 32B-class coder? M5 Pro with the 20-core GPU and 64GB (14-inch from $2,699 at 24GB, plus the memory upgrade) holds 32B-class models with room for context and saves well over a thousand dollars.
4. 8B models and curiosity? Any base M5 or the new M6. Even 16GB runs an 8B at speed. Spec 32GB if you can and check the 32GB Mac model picks.
5. Own an M4 Max 128GB? Skip this generation unless long prompts are your bottleneck; the prefill gain is the real upgrade, and it rarely justifies $4K+.
6. Need raw speed on models under 32GB? A GPU tower still wins. Different tool, different job.

The memory shortage briefly turned a laptop into Apple's flagship AI machine by subtraction. With the M5 Ultra Mac Studio shipping, the MacBook Pro M5 Max 128GB is back to being what it should be: the most capable local-AI machine you can carry.

---

Sources

• Apple Newsroom — MacBook Pro with M5 Pro and M5 Max — March 3, 2026 announcement: original prices, memory tiers, bandwidth, AI claims
• Apple Newsroom — Mac Studio with M5 Max and M5 Ultra — August 25, 2026: up to 512GB, 1.2TB/s, availability September 22
• Apple Newsroom — 14-inch MacBook Pro with M6 — October 27, 2026: M6, up to 32GB and 170GB/s
• Apple Newsroom — MacBook Pro with PROLED XDR Display — October 28, 2026: M5 Pro and M5 Max models with a PROLED display, from $2,899, shipping November 6
• Apple — MacBook Pro tech specs and Mac Studio tech specs — current chip, memory, and bandwidth configurations (read November 2, 2026)
• Apple Store — Buy MacBook Pro and Buy Mac Studio — US starting prices (read November 2, 2026)
• llama.cpp Apple Silicon performance thread — community llama-bench results per chip, LLaMA 7B at F16, Q8_0 and Q4_0 (read November 2, 2026)
• tbreak — Mac Studio M5 Max review — September 21, 2026: MLX and Ollama generation speeds on a 40-core, 128GB M5 Max
• MacRumors — Mac Studio roundup — the 512GB/256GB/128GB M3 Ultra memory-tier removals amid rising 2026 RAM prices

---

FAQ