Apple M1 vs M4 for local AI: memory, bandwidth, and what actually fits

I'm Matt Kerr, and I build Outlier, a Mac app that runs language models on your own hardware — so I'm an interested party, worth saying up front. This isn't a pitch. It's the hardware note I wish I'd had before spending two years profiling inference on these chips, and the honest version includes where a PC with a discrete GPU wins.

Unified memory is why a Mac is in this conversation

On a typical PC a model has to fit the graphics card's VRAM: 12GB card, 12GB budget, and the system RAM sitting next to it doesn't help. Apple silicon shares one pool between CPU and GPU, which is why MLX, Apple's MIT-licensed array framework, is built on the line "Arrays in MLX live in shared memory." No copy across a PCIe bus, no VRAM cliff.

M1 has this too. Unified memory shipped in November 2020 and hasn't changed shape since, so a 16GB M1 holds a model that would otherwise need a 16GB card. The advantage is architectural, not generational.

Capacity and bandwidth are two different limits

Capacity is a wall — the weights fit in RAM or the machine swaps to SSD and the experience collapses. Bandwidth is a slope: generating each token means reading the weights out of memory, so tokens per second scales close to linearly with read speed.

The numbers line up almost too neatly. Apple never published a bandwidth figure for the base M1, but called M1 Max's 400GB/s "nearly 6x that of M1," and later said M5's 153GB/s is "more than 2x over M1" — both imply about 68GB/s. M4 is a documented 120GB/s, a 1.76x ratio. In llama.cpp's community table (LLaMA 7B, Q4_0), M1 generates 14.15 tok/s and base M4 generates 24.11 — 1.70x. Decode is a bandwidth story, and it behaves predictably.

Prefill is the other half, and it's compute-bound, so GPU core count shows up: 117.96 tok/s on M1, 221.29 on M4, 439.78 on M4 Pro, 885.68 on M4 Max. Paste in a long document and prefill is the wait before the first word appears.

What each configuration actually fits

Same question, concrete. I'm using Outlier's tiers here because those RAM minimums are measured rather than guessed.

Two caveats. 16GB is the M1 maximum — Apple's spec sheet for the M1 MacBook Air reads "8GB unified memory, Configurable to 16GB," and plenty shipped with 8GB. LM Studio still tells 8GB owners to "stick to smaller models and modest context sizes." Chip ceilings aren't machine ceilings either: Apple's spec page tops the M4 Mac mini at 24GB, not 32GB.

The pool is also shared. macOS, your browser and the app itself draw from the same RAM the weights need, so usable budget sits below the nameplate. Metal exposes a "recommended max working set size" for exactly this; I couldn't retrieve Apple's documented fraction, so I won't quote one.

The Neural Engine numbers are the wrong number

Most M1-vs-M4 pieces lead with 11 TOPS versus 38. For local LLMs that's inert. MLX lead maintainer Awni Hannun wrote in the project's repo that "at the moment we don't have plans to support ANE in MLX given it is a closed source API," reconfirmed it in 2025, and the request is closed as wontfix. llama.cpp runs on Metal, and Ollama's blog says its Apple silicon build now sits on MLX. Your tokens come off the GPU on both chips.

There's a second problem: M4's 38 TOPS is widely reported as an INT8 measurement against M1's FP16 figure. I found no Apple page stating either precision, so treat cross-generation TOPS as not directly comparable.

Sustained load, and what I won't estimate

Local inference is sustained, not bursty — a long generation pins the GPU for its whole duration. On thermals, the honest answer is that I have no vendor-published sustained-inference data comparing a fanless Air to an actively cooled mini, and I'm not inventing a table. Architecturally, though, both chips do this at laptop power, quietly, on battery, which no discrete-GPU laptop matches. If you'll run hours of generation a day, pick the actively cooled chassis whatever chip is inside it.

Price per usable gigabyte, and the upgrade path that doesn't exist

Unified memory is soldered. Neither chip has an upgrade path, which makes the configure-to-order choice the most consequential decision you'll make for local AI. Buy memory, not clock speed.

The prices Apple publishes: Mac mini with M4 and 16GB at $599, Mac mini with M4 Pro at $1,399. That's roughly $37 per gigabyte of unified memory at the entry point, and it's the cheapest new Mac that runs mid-size models decently. M1 machines aren't sold new at all — the MacBook Air line is M5 only now — so an M1 is a used or refurbished buy, and I have no sourced figure for street prices, so check rather than trust an article.

One currency note: M4 isn't Apple's current silicon. M5 shipped in October 2025 with a Neural Accelerator in every GPU core and 153GB/s bandwidth, and Apple claims "over 4x the peak GPU compute performance for AI compared to M4." M5 Pro and M5 Max followed in March 2026.

Where a discrete GPU still wins

On raw throughput, a PC with a good discrete GPU is usually faster than a Mac at the same price, and I'd rather say that than pretend otherwise. Dedicated graphics memory runs at bandwidths no base Apple chip approaches, and the CUDA ecosystem for training, fine-tuning and kernel work is deeper than the Metal side. If your model fits comfortably in VRAM, buy the card.

Apple's case is capacity per dollar and power draw. A discrete card makes you shrink the model to fit; unified memory lets you scale the model to the machine — Apple markets the M4 Max's 128GB as enough to "easily interact with large language models that have nearly 200 billion parameters," and the M3 Ultra Mac Studio configures to 512GB. Matching that capacity in cards costs far more and pulls far more power.

Where Apple silicon genuinely wins — M1 included

The M1 isn't a legacy device, and people write it off too fast. Apple Intelligence still requires only "M1 and later." LM Studio's stated minimum is Apple silicon plus macOS 14. Apple's Foundation Models framework — a ~3B on-device model, offline, with "AI inference that is free of cost" — runs on any Apple Intelligence-capable Mac. And 14 tok/s on a 7B model is faster than most people read.

What M4 buys is the ceiling, not the speedometer. Going from 16GB to 32GB (or 64, or 128) moves you out of "7-8B only" and into 13B-32B territory — a category change, not a percentage. In Outlier's tiers that's Lite 9B versus Core 27B, the tier that matched Claude Opus on 98.9% of rubric checks across a 54-prompt head-to-head. Model size decides whether a local app feels useful far more than tokens per second does.

Who should pick which

Already own a 16GB M1: keep it. Run 4B-9B models with sane contexts and you have a working local setup today — on Outlier that's the free Nano and Lite tiers, no account.

Own an 8GB M1: capacity is your wall and tuning won't move it. Nano (2.37GB download, 6GB minimum) works; a 27B model never will.

Buying new on a budget: the $599 M4 Mac mini is a fair entry point, but spend the configure-to-order money on memory first — and compare current M5 machines before you commit.

Serious about big models: you want M4 Pro or Max with 64GB or more, so $1,399 and up. Outlier's Plus tier (397B-a17b, 209GB download) needs 64GB and streams experts from SSD to hold peak RSS near 11GB; it runs at 2.1 tok/s on my M1 Ultra — useful for hard problems, nobody's idea of chat.

Throughput above all, on one model that fits a card: buy the PC. I'd rather you got the right machine than the one I write software for.