Apple's M1 and M2 Macs Are Falling Behind for Newer AI Models

The M1 and M2 Macs earned their reputation honestly. They were fast, cool, and quiet, and for everyday work — browsing, writing, video calls, even serious photo and video editing — an M1 MacBook Air still holds up years later. But one fast-moving workload is leaving those machines behind, and it happens to be the one everyone suddenly cares about: running newer AI models locally. The problem isn't that the M1 and M2 chips got slow. It's that the models got big, and the single thing you can never change on an Apple Silicon Mac is how much memory it has.

The models grew; the memory didn't

A local large language model has to fit entirely in a Mac's unified memory to run well. That was easy when capable open models were small. It is getting hard fast. The current Qwen3 generation (which replaced Qwen2.5) and its peers now range from compact models up to variants with tens or even hundreds of billions of parameters — and the genuinely useful mid-sized versions increasingly want more RAM than an entry-level Mac has to give.

That is where the M1 and M2 hit a wall. The base M1 shipped with just 8GB of memory and topped out at 16GB. The M2 improved things only modestly: as Apple noted at launch, it offered "100GB/s of unified memory bandwidth" and "up to 24GB of fast unified memory." Respectable numbers in 2022 — but because Apple Silicon memory is soldered to the chip and fixed at purchase, an 8GB or 16GB machine is stuck there for life. You cannot add a stick of RAM to catch up.

Two limits, not one: capacity and bandwidth

Newer models strain M1 and M2 Macs in two different ways.

Capacity is the hard stop. If a model plus its working memory is larger than the RAM available, it simply won't load — or it spills over and slows to a crawl. An 8GB Mac has room for only the smallest, heavily compressed models once the operating system takes its share. A 16GB Mac does better, but still can't touch the mid-sized models that make local AI genuinely useful.

Bandwidth sets the speed. Every token a model generates requires reading through its weights, so memory bandwidth largely determines how quickly it responds. The M1's roughly 68GB/s and the M2's 100GB/s were fine for their era, but Apple has since multiplied that figure several times over — leaving the earliest chips generating text noticeably slower on anything demanding.

The gap to M3, M4, M5, and M6 is widening

Each generation since has pushed memory further, and Apple is now designing explicitly around on-device AI. In announcing its newest silicon, Apple said the M6 "supports up to 32GB of unified memory ... [to] run LLMs on device," while the top-end M5 Ultra scales to 512GB and can "run huge LLMs with hundreds of billions of parameters entirely on device." Against a 512GB ceiling — or even a routine 32GB one — an 8GB M1 is in a different category entirely for this particular job. For general computing the gap is modest; for running newer models it is a chasm.

What an M1 or M2 still does well

None of this is a reason to panic or to write off a perfectly good Mac. If your local-AI needs are modest, an M1 or M2 remains capable:

• Small, quantized models in the roughly 8B range still run on a 16GB machine, which is plenty for chat, summarizing, and light coding help.
• The tooling still supports these chips well — llama.cpp calls "Apple silicon ... a first-class citizen," and Apple's own MLX framework runs across every M-series generation.
• For everything that isn't a large language model, an M1 or M2 is still a genuinely fast computer.

The dividing line is specific: it's the newer, larger models — not computing in general — that these machines can no longer keep up with.

The upgrade math, and what to do with the old Mac

If running current models locally matters to you, the uncomfortable truth is that there is no software fix. You can't add memory to an M1 or M2, so keeping pace means moving to a higher-RAM M3, M4, M5, or M6 machine. The good news is that the same forces stranding your old Mac for AI are propping up its resale value: capable Apple Silicon Macs are in demand, their memory can't be upgraded, so used high-RAM units are genuinely scarce — and that helps an older Mac hold its price better than most technology does.

That makes the cleanest path an upgrade-and-sell: put the value of the machine you're replacing toward the one that can actually run what you want. If you go that route, see what your current Mac is worth before you buy — the offer is what you're paid, with no renegotiation once it arrives, and it takes about 30 seconds to find out.

The bottom line

The M1 and M2 didn't fail; the workload changed underneath them. Local AI models have grown past what 8GB and 16GB of fixed, moderate-bandwidth memory can comfortably hold, and each new Apple Silicon generation widens the gap for that one task. For everyday use, hang on to a good M1 or M2 with confidence. But if your interest is running the newest models at home, it may be time to upgrade — and to let the old machine help pay for the new one.