← All posts

M6 Mac mini for Local LLMs: Is the Upgrade Worth It?

Apple just dropped the M6, the M5 Ultra, and the M5 Pro, and the local-LLM corner of the internet has gone predictably feral. Every YouTube thumbnail is a shocked face next to a bar chart. Every Reddit thread is someone asking whether they should sell their M4 Mac mini right now to fund an upgrade.

I've been running local models on a Mac mini for the better part of a year — I wrote the practical guide back in June — and I've watched this exact cycle play out with every single Apple Silicon release. The pattern is always the same: a spec sheet full of impressive numbers, a wave of "this changes everything" takes, and then a quiet reckoning when people realize the number that actually matters for inference didn't move as much as the marketing implied.

So let's skip the spec recap. You can read the press release anywhere. The question that matters — the only question, if you're running LLMs locally — is this: does the new chip move tokens faster, and is that speed worth the price of admission?

The answer is more nuanced than the hype suggests, and it comes down to one number most reviewers barely mention: memory bandwidth.

The Only Number That Matters: Memory Bandwidth

Here's the thing about local LLM inference that most "should you upgrade" articles gloss over: you are not compute-bound. You are memory-bandwidth-bound.

When you run a model locally, the weights live in unified memory. Every single token you generate requires reading essentially the entire model's weights from memory. A 30B-parameter model at 4-bit quantization is roughly 16GB of weights. To generate one token, the chip has to stream all 16GB of those weights through its compute units.

That means your tokens-per-second is, to a first approximation, capped by a brutally simple formula:

tokens/sec ≈ memory_bandwidth / model_size_in_memory

If your chip can read 400 GB/s and your model is 16GB, you're looking at a theoretical ceiling of about 25 tokens/sec — and in practice you'll get maybe 60-70% of that after overhead. If your chip reads 800 GB/s, the same model roughly doubles.

Notice what's not in that formula: CPU core count. GPU core count. Neural Engine TOPS. Clock speed. None of it matters for the steady-state speed of text generation, because none of it changes how fast you can pull weights out of memory. A chip with twice the compute and the same memory bandwidth will generate tokens at almost exactly the same speed.

This is why the M-series memory bandwidth numbers are the single most important spec for local LLM users — and why Apple's own marketing, which leads with CPU and GPU core counts, is actively misleading for this use case.

The M-Series Bandwidth Ladder

Let's lay out the numbers that actually matter, because this is where the upgrade decision lives or dies.

Chip Memory Bandwidth Typical Mac mini Config
M1 ~68 GB/s 8-16GB
M2 ~100 GB/s 8-24GB
M2 Pro ~200 GB/s 16-32GB
M3 ~100 GB/s 8-24GB
M4 ~120 GB/s 16-32GB
M4 Pro ~273 GB/s 24-64GB
M5 ~150 GB/s (est.) 16-32GB
M5 Pro ~300 GB/s (est.) 24-64GB
M6 ~180 GB/s (est.) 16-48GB
M5 Ultra ~800+ GB/s (Mac Studio, not mini)

(The M5/M6 figures are based on launch-window reporting and Apple's own disclosures; I've flagged estimates where the exact number hasn't been independently confirmed. The point stands regardless of the precise figure.)

Look at that ladder carefully, because it tells you almost everything you need to know.

The jump from M4 to M6 on the base Mac mini is roughly 120 GB/s to 180 GB/s — a 50% improvement. That's real, and it's not nothing. But it's also not the "2x faster" that the launch headlines imply, because the headlines are comparing CPU and GPU core counts, not memory bandwidth.

The genuinely big jump is the M5 Ultra at 800+ GB/s — but that's a Mac Studio part, not a Mac mini, and it starts at a price point where you have to ask yourself whether you're buying a computer or a down payment on a used car.

What 50% More Bandwidth Actually Buys You

Let's make this concrete, because "50% more bandwidth" is an abstraction until you attach it to a model you'd actually run.

Say you're running a 14B model at 4-bit quantization — roughly 8GB of weights. On an M4 Mac mini (120 GB/s), you're looking at a theoretical ceiling of about 15 tokens/sec, and a realistic 9-11 tokens/sec in practice. That's usable — it's roughly the speed of a fast human typist — but it's not snappy.

On an M6 Mac mini (180 GB/s), the same model's ceiling jumps to about 22 tokens/sec, realistic 14-16 tokens/sec. That's the difference between "I can tolerate this" and "this feels responsive." For interactive use — coding assistance, chat, brainstorming — that jump is genuinely meaningful.

But here's the catch: the models you can actually fit in a Mac mini's memory haven't changed. The M6 base mini tops out at 48GB of unified memory, and the M5 Pro at 64GB. That's the same ceiling the M4 Pro already hit. You're not suddenly running 70B models on a Mac mini because of the M6 — you're running the same models, just a bit faster.

And that's the crux of the upgrade question. The M6 doesn't unlock a new tier of model. It makes the tier you already have access to run somewhat faster.

The Real Upgrade Question, Reframed

So should you upgrade? Let me give you the honest framework I'd use, because "it depends" is a cop-out.

Upgrade if you're on M1 or M2. The jump from ~68-100 GB/s to ~180 GB/s is a 2-3x improvement in memory bandwidth, and that is transformative. A 14B model goes from "painfully slow" to "actually usable." If you've been limping along on an M1 mini running 7B models and wishing you could step up, the M6 is a legitimate, defensible upgrade. This is the one case where the hype is roughly justified.

Upgrade if you're memory-constrained, not speed-constrained. If your current mini has 16GB and you're constantly hitting the ceiling — swapping, running quantized-to-death models, unable to load the 14B you want — then moving to a 32GB or 48GB M6 is worth it. But understand that what you're buying is memory capacity, not the M6 specifically. An M4 Pro with 64GB would serve you just as well for most workloads, and it's cheaper now that it's "last gen."

Don't upgrade if you're on M4 or M4 Pro. This is the uncomfortable truth the launch-window hype doesn't want you to hear. If you bought an M4 or M4 Pro Mac mini in the last year, the M6 is a marginal improvement for local LLMs. You're looking at maybe 30-50% more tokens/sec on the same models, for the cost of a whole new machine. That's not nothing, but it's also not the kind of leap that justifies selling a year-old computer at a loss.

Don't upgrade if you want to run bigger models. The M6 Mac mini does not meaningfully change the model-size ceiling. If your goal is to run 70B-class models locally, a Mac mini was never the right tool, and the M6 doesn't change that. You want a Mac Studio with an Ultra chip, or — more sensibly for most people — a used server with a couple of 24GB GPUs. The M6 mini is a faster small-model machine, not a bigger-model machine.

The Bandwidth-to-Dollars Math

Let me put a number on it, because that's what actually settles these debates.

The base M6 Mac mini is expected to land around $599-699. The M4 Mac mini it replaces launched at $599. So you're paying roughly the same price for roughly 50% more memory bandwidth. On a per-token-per-dollar basis, that's a genuinely good deal if you're coming from an older machine.

But if you already own an M4, the calculus flips. Selling your M4 mini gets you maybe $350-400 on the used market. Buying an M6 costs $599-699. So your net cost to go from 120 GB/s to 180 GB/s is $200-350 — for a 50% speed bump on models you can already run. That's a hard sell unless you're doing local inference constantly and the speed genuinely frustrates you.

Compare that to the M5 Ultra in a Mac Studio, which starts around $3,999. That's where the 800+ GB/s lives, and it's the only chip in this launch that meaningfully changes what you can run. But at four grand, you have to be serious about local inference — and at that point, you should also be pricing out a used dual-3090 rig, which will run circles around it for less money if you don't mind the power draw and the noise.

What I'd Actually Do

Here's my honest recommendation, as someone who runs local models on a Mac mini every day and has no particular loyalty to Apple's release cadence.

If you're on M1/M2: Upgrade. The M6 is the first base Mac mini where a 14B model is genuinely pleasant to use, and the price is right. This is the upgrade that makes sense.

If you're on M4: Keep it. The M6 is faster, but not fast enough to justify the cost of a year-old machine. Wait for the M7, or — better — wait until the M5 Pro Mac minis hit the used market in a year and grab one of those for the memory capacity bump.

If you're memory-starved: Buy memory, not chip. A 64GB M4 Pro or M5 Pro is a better use of your money than a 32GB M6, because for local LLMs, capacity beats bandwidth once you're past the "usable speed" threshold. A model that fits in memory and runs at 10 tokens/sec beats a model that doesn't fit at all.

If you want big models: Stop looking at Mac minis. The M6 doesn't change the ceiling. Look at a Mac Studio Ultra, or a used GPU server, or — honestly — just use a cloud API for the occasional 70B task and keep your mini for the everyday 7-14B work.

The Bottom Line

The M6 Mac mini is a good chip. It's a meaningful step forward for local LLM users coming from M1 or M2, and it's a marginal step forward for everyone else. The launch-window hype is doing what launch-window hype always does: leading with the numbers that sound impressive (core counts, TOPS, "2x faster") and burying the number that actually matters for inference (memory bandwidth, up ~50% on the base mini).

The honest summary is this: the M6 makes the models you can already run faster. It does not make bigger models runnable. If that's what you need, it's a good buy. If you were hoping for a local-LLM revolution, you'll be waiting for the next Ultra — or for someone to finally ship a Mac mini with the memory bandwidth of a Mac Studio.

Until then, the best local-LLM machine is still the one you already own, running the right model for its memory, at a speed you can live with. Everything else is just a bar chart.


Are you upgrading for local LLMs, or sitting this one out? I'd love to hear your bandwidth-to-dollars math — find me on GitHub or drop a note in the comments.

If you found this useful, you might also like my practical guide to running local LLMs on a Mac mini, the local LLM homelab hardware guide, my LocalAI vs Ollama comparison, and the full gear list of everything I actually run.