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 is the hardware note I wish I'd had, and it includes where a PC with a discrete GPU wins.

Unified memory is why a Mac is in this conversation

On a PC, a model has to fit the graphics card's VRAM: 12GB card, 12GB budget, and the system RAM 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 premise that its arrays live in shared memory.

The M1 has this too. Unified memory shipped in November 2020, 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 falls back to disk and the experience collapses. Bandwidth is a slope: each token means reading the weights out of memory, so tokens per second scales with read speed.

Apple never published a bandwidth figure for the base M1, but it called M1 Max's 400GB/s nearly six times M1's, and later put M5's 153GB/s at more than double M1's. 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 manages 24.11, a 1.70x gap. Decode is a bandwidth story.

Prefill is compute-bound, so GPU cores show up there instead: 117.96 tok/s on M1 against 885.68 on M4 Max. That's the wait before the first word appears.

M1 Pro and M1 Max: more bandwidth than a base M4

The base M1 is the bottom of its family. Apple's release for the M1 Pro and M1 Max states up to 200GB/s with up to 32GB of unified memory for the M1 Pro, and up to 400GB/s with up to 64GB for the M1 Max. Set against the M4 figures on this page, an M1 Pro (200GB/s) sits above a base M4 (120GB/s) and below an M4 Pro (273GB/s), and an M1 Max (400GB/s) sits above an M4 Pro, with only the M4 Max (up to 546GB/s) ahead of it.

Because generation speed follows bandwidth, those numbers predict that an M1 Pro or M1 Max generates faster than a base M4. I have not measured either, and this page has no benchmark row for them, so read that as what the specs imply, not as a result. Capacity is the other limit: an M1 Pro tops out at 32GB, the same ceiling as a base M4, and prefill depends on GPU cores, where the newer chips pull ahead.

What each configuration actually fits

The "what fits" row uses Outlier's tiers because those RAM minimums are measured rather than guessed.

Two caveats. 16GB is the M1 maximum, not its typical configuration. Apple's spec sheet for the M1 MacBook Air reads "8GB unified memory, Configurable to 16GB," which is why LM Studio tells 8GB owners to stick to smaller models. Chip ceilings aren't machine ceilings either, and the M4 machines are gone from Apple's buy pages: the Mac mini is now M6 and M5 Pro, the Mac Studio M5 Max and M5 Ultra, the MacBook Air M5, and the MacBook Pro M5, M5 Pro and M5 Max. So the 64GB and 128GB in that row are what the M4 silicon supports, not what you can order. On Apple's spec pages today, 64GB is the ceiling on the M5 Pro Mac mini and 128GB on the M5 Max Mac Studio and MacBook Pro. And the throughput figures are community-submitted, one model at one quantization.

The pool is shared, too: macOS and your browser draw from the same RAM the weights need, so usable budget sits below the nameplate. Metal exposes a "recommended max working set size" for this, but 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 comparisons lead with 11 TOPS against 38. For local language models that figure is inert. MLX's lead maintainer wrote in the project's own repo that there are no plans to support the ANE "given it is a closed source API," and the request is closed as wontfix. Your tokens come off the GPU on both chips.

M4's 38 TOPS is also 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.

Thermals, price per gigabyte, and the upgrade path that isn't

Local inference is a sustained load, not a burst. A long generation pins the GPU for its whole duration. I have no vendor-published data comparing a fanless Air against an actively cooled mini under that load, so I won't put numbers on it. If you'll generate for hours a day, pick the actively cooled chassis. Both chips do this at laptop power, quietly, on battery, a combination few discrete-GPU laptops sustain at comparable power draw.

Unified memory is part of the chip package and chosen when you order. Apple's spec language is "configurable to." There's no aftermarket upgrade, so build-to-order is the most consequential decision here. Buy memory first.

Apple no longer sells an M4 Mac mini, so I have no current M4 price to quote. Apple's Mac mini buy page, read 2026-10-19, starts at $899 on M6 and $1,699 on M5 Pro (apple.com/shop/buy-mac/mac-mini), so an M4 is a used or refurbished buy at a price I can't source. M1 machines aren't sold new any more either (the Air line is M5 only), so the same goes for an M1. Which raises the currency problem: M5 shipped in October 2025 at 153GB/s, with Apple claiming over 4x M4's peak GPU compute (and over 6x M1's peak GPU compute for AI).

Where a discrete GPU wins, and where Apple silicon does

On raw throughput, a PC with a good discrete GPU is usually faster than a Mac at a similar price, and I'd rather say so than pretend otherwise. Dedicated graphics memory runs at bandwidths no base Apple chip approaches, and CUDA's tooling is deeper than Metal's. If your model fits comfortably in VRAM, buy the card.

Apple's counter isn't speed, it's capacity per dollar and watts per token. A card makes you shrink the model to fit it; unified memory lets you scale the model to the machine. Apple markets M4 Max's 128GB as enough for models of nearly 200 billion parameters. Matching that in cards costs far more and draws far more power.

The M1 gets written off too fast, too. Apple Intelligence still requires only "M1 and later," and LM Studio's minimum is Apple silicon on macOS 14. Apple's Foundation Models framework hands third-party apps a roughly 3B on-device model, free and offline, on any Apple Intelligence-capable Mac, though the weights aren't published. And 14 tok/s on a 7B model is faster than most people read.

What M4 buys is the ceiling, not the speedometer: 16GB to 32GB moves you out of "7-9B only" and into 27B-class territory. That's a category change, not a percentage. In Outlier's tiers that's Lite 9B versus Core 27B. We published a 54-prompt side-by-side of Core 27B and Claude Opus 4.7 so you can compare the answers yourself; Claude is still faster and stronger on the hardest work.

Pick the chip by the model you want to run

You have an 8GB M1: capacity is your wall and tuning won't move it. A 4B model works well (Outlier's Nano is a 2.37GB download, 6GB minimum), and a 27B never will.

You have a 16GB M1: keep it. 4B to 9B models with sensible context lengths are a working local setup today. Upgrade when you want 27B-class quality.

Buying new on a budget: Apple's entry Mac mini is now the $899 M6, so put the build-to-order money into memory and compare the current M5 machines first. An M4 is a used buy.

Long documents or codebases: pay for GPU cores. Prefill is where M4 Pro and Max earn their price.

The largest models: 64GB minimum, and Apple sells no M4 machine that gets there. Today that means an M5 Pro Mac mini configured to 64GB, a Mac Studio (the M5 Max starts at $2,499 with 36GB, and Apple lists 64GB and 128GB as build-to-order options on the 40-core-GPU M5 Max), or a MacBook Pro with M5 Max at 128GB. Outlier's Plus tier (397B-a17b, 209GB) pages experts off the SSD to hold peak RSS near 14GB: 1.59 tok/s on my M1 Ultra, useful for hard problems and nobody's idea of chat.

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