TL;DR: The M6 Mac Mini (preorders open now, ships September 22) tops out at 32GB of unified memory at 170 GB/s — not the 200 GB/s the June rumors promised. The $1,299 32GB config fits Qwen3.8-27B with room to spare and runs silently at a fraction of GPU power draw, but a used RTX 3090 is 4–5× faster on the same models. Buy it as a quiet always-on AI appliance, not as a speed machine.

	Mac Mini M6 32GB	Used RTX 3090 PC	Mac Mini M5 Pro
Best for	Silent always-on home AI server	Fastest tokens per dollar	Bigger models on a Mac
Memory for models	32GB @ 170 GB/s	24GB @ 936 GB/s	up to 64GB @ 307 GB/s
Price	$1,299	~$1,000–$1,350 (card only)	from $1,699
The catch	~8–10 tok/s on dense 27B (est.)	285–350W, loud, used silicon	Apple RAM pricing on the way up

Honest take: if you want the fastest local AI under $1,500, buy a used RTX 3090 — it’s 4–5× the speed on every model both machines can hold. Buy the M6 Mini 32GB only if silence, 10–20W idle power, and a warranty matter more to you than tokens per second — and if you’re spending Mac money for AI, the $1,699 M5 Pro with its 307 GB/s is the better Mac.

Apple announced the new Mac Mini M6 on August 25, 2026, alongside the M5 Ultra Mac Studio, and machines reach desks September 22. The Reddit and Hacker News framing — “200 GB/s unified memory, best sub-$1K local LLM machine?” — has been circulating since June. Now that Apple has published the actual spec sheet, both halves of that framing need correcting. Here’s what the M6 Mini actually is, what it runs, and who should buy one instead of a GPU.

What Apple actually shipped (vs. what the rumors said)

Back in June, Bloomberg’s Mark Gurman reported that the base M6 would push unified memory bandwidth to roughly 200 GB/s, up from 153 GB/s on the M5 base — we covered that report in our M7 roadmap analysis and flagged the number as the only bandwidth figure in circulation. That number did not survive contact with the spec sheet.

What shipped, per Apple’s announcement and the launch coverage (Apple Newsroom, MacRumors, AppleInsider):

Config	Memory	Bandwidth	Price
M6 base	16GB / 256GB SSD	153 GB/s	$899
M6 mid	24GB	170 GB/s	$1,099
M6 top	32GB	170 GB/s	$1,299
M5 Pro	up to 64GB	307 GB/s	from $1,699

The M6 itself is a genuinely new chip — Apple’s first on 2nm, with a 12-core CPU (two new “super cores,” four performance, six efficiency), a 12-core GPU with a Neural Accelerator in every core, and a dual 16-core Neural Engine that Apple says doubles peak Neural Engine compute. Apple claims the fastest single-threaded performance of any consumer chip. None of that is marketing fluff for general computing.

But for local LLM inference, one number dominates everything else on the sheet: memory bandwidth. Token generation is memory-bandwidth-bound — the GPU reads every active parameter for every token — so 153–170 GB/s is the speed limit, and no amount of Neural Engine compute changes it. We made exactly this argument about the M7’s “AI-focused” branding, and it applies verbatim to the M6’s dual Neural Engine: great for Apple Intelligence features, nearly irrelevant to llama.cpp and MLX decode speed.

The pricing also deserves a correction. “Sub-$1K local LLM machine” describes only the $899 16GB config — and 16GB of unified memory is not a local-LLM config in 2026. macOS reserves memory for itself: by default, Metal lets the GPU wire roughly 75% of unified memory, so a 16GB machine offers about 12GB to models. That runs 7B–8B models well and 12–14B models tightly, and nothing from the 27B class that makes this year’s local models interesting. The machine this article is actually about — the one that fits Qwen3.8-27B — costs $1,299.

What fits in 32GB

The 75% rule gives the 32GB M6 roughly 24GB of GPU-usable memory out of the box — coincidentally, the same working budget as an RTX 3090, with the difference that you can raise the Mac’s ceiling. Check and raise it like this:

$ sysctl iogpu.wired_limit_mb
iogpu.wired_limit_mb: 0        # 0 = macOS default, ~75% of RAM

$ sudo sysctl iogpu.wired_limit_mb=28672   # allow 28GB of 32GB (resets on reboot)

Leave at least 4GB for macOS — pushing the limit to the full 32GB invites memory pressure and swap-driven stalls (ModelPiper’s wired-limit guide covers the failure modes).

With that budget, here’s the fit table for the models a 2026 buyer actually cares about:

Model	4-bit size	Fits 32GB M6?	Fits 16GB M6?
Qwen3.8-27B Q4_K_M	16.8GB	Yes, ~7GB context headroom	No
Qwen3.6-35B-A3B Q4	~21GB	Yes (raise wired limit for long context)	No
IFM K2 Horizon 32B 4-bit	~18–20GB	Yes	No
Gemma 4 26B-A4B QAT	~15GB	Yes, comfortable	No
7B–8B class Q4	4–5GB	Yes, trivially	Yes
Dense 70B Q4	~40GB	No	No

That top row is the purchase justification. Qwen3.8-27B — the best open model that fits consumer hardware — needs 17–19GB of VRAM or unified memory per Unsloth’s guidance, which excludes every 16GB card and every 16GB Mac. The 32GB M6 Mini is now the cheapest new Apple hardware that clears the bar, and it clears it with context room to spare. The 32GB ceiling also happens to be exactly where Ollama v0.30 turns on MLX by default on Apple Silicon — the 16GB and 24GB configs miss that threshold too.

Run the numbers for your own model list with our VRAM calculator — unified memory entries work the same way as VRAM, minus the ~25% macOS reserve.

What speed to expect (bandwidth math, honestly labeled)

No M6 Mini has shipped, so no measured tok/s exists yet — every number in this section is bandwidth-proportional extrapolation, the same method that called the M5 Max within its measured range. It works because Apple Silicon decode speed scales almost linearly with memory bandwidth: the M4 Pro (273 GB/s) measures 49.64 tok/s on a 7B Q4_0 in llama.cpp, and the M4 base (120 GB/s) measures 24.11 tok/s on the identical benchmark — a 2.06× speed gap on a 2.28× bandwidth gap (llama.cpp Apple Silicon thread; full breakdown in our M4 Pro Mini guide).

Scaling from the M4 Pro’s measured numbers, at 170 GB/s (62% of the M4 Pro’s bandwidth):

Model	M6 32GB (est.)	M4 Pro 24GB (measured/community)	Used RTX 3090 (measured)
7B–8B Q4	~28–33 tok/s	49.6 tok/s	~95 tok/s
14B Q4	~14–16 tok/s	22–25 tok/s	~55 tok/s
Qwen3.8-27B Q4_K_M	~8–10 tok/s	(doesn’t fit 24GB Mac usable)	41 tok/s
Qwen3.6-35B-A3B (MoE, 3B active)	~45–65 tok/s	—	107 tok/s

MLX typically adds 20–30% over llama.cpp’s Metal backend on the same hardware, so treat the upper ends as the MLX numbers. The MoE row is the one to internalize: sparse models read only ~3B parameters per token, so Qwen3.6-35B-A3B should decode faster on this 170 GB/s machine than a dense 14B does. If you buy an M6 Mini for local AI, MoE models are your daily drivers; dense 27B at 8–10 tok/s is usable for chat (reading speed is ~7–10 tok/s) but painful for agentic loops.

Setup is two commands once you’re on the 32GB config: ollama pull the model and check ollama ps shows 100% GPU, or install MLX directly with pip install mlx-lm and point mlx_lm.generate at a 4-bit community conversion from the mlx-community org on Hugging Face. Prompt processing remains Apple Silicon’s weak leg — prefill is compute-bound, and long-context prefills that take seconds on a discrete NVIDIA card take tens of seconds on M-series, dual Neural Engine or not.

M6 Mini 32GB vs. used RTX 3090: the actual decision

This is the comparison the Reddit thread was really asking about, and at September 2026 prices it’s closer than the raw specs suggest — on cost, not on speed.

A used RTX 3090 runs $1,000–$1,350 depending on tracker and week (BestValueGPU has September asks near the low end; ResalePrices’ market average is $1,343, up 11% in 90 days as the DRAM crisis drags used cards up). That’s the card alone: a complete quiet-enough tower around it lands at $1,600+ in the current RAM market, against $1,299 for the Mac that’s a complete computer out of the box.

What each dollar buys:

• Speed: the 3090’s 936 GB/s is 5.5× the M6’s bandwidth, and the measured numbers track it — 41 vs ~8–10 tok/s on Qwen3.8-27B, 107 vs ~45–65 on the 35B MoE. Every model both machines fit, the 3090 runs 4–5× faster. It also front-loads prompts an order of magnitude faster, which matters for coding assistants and RAG far more than decode speed does.
• Power and noise: the 3090 pulls 285–350W under inference load — $0.034/hour at $0.12/kWh, audible from the next room. Apple rates the entire new Mini chassis at 155W maximum continuous, and the outgoing M4-generation minis measured 4–5W idle and 65–140W at absolute peak (Apple’s power spec sheet); an M6 doing LLM inference will sit in the tens of watts. For an always-on server, the electricity delta alone is $15–25/month.
• Capacity: 32GB of unified memory (28GB usable, raised) vs 24GB of VRAM. The Mac fits Q5/Q6 quants of the 27B class and longer contexts that spill past 24GB. Neither machine touches dense 70B — for that you need a Max-tier Mac or partial offload.
• Risk: four-to-five-year-old used silicon with no warranty vs a new machine with one.

And there’s an in-family problem for the M6: at $1,699, the M5 Pro Mini’s 307 GB/s is 1.8× the M6’s bandwidth with a 64GB ceiling (Macworld’s launch coverage). Configured to 48–64GB it climbs past $2,000 and starts competing with the Mac Studio M5 Max — but at the $1,699 base it makes the $1,299 M6 look like the wrong Mac for anyone whose main workload is inference. If Apple money is already committed, the extra $400 buys nearly double the tokens per second.

What to actually buy

Prices as of September 2026, all taken from the comparison above:

Your situation	The machine	Price	Where
Fastest local AI under $1,500, noise be damned	Used RTX 3090	~$1,000–$1,350	Check price
Silent always-on box for MoE models + home automation	Mac Mini M6 32GB	$1,299	Check price
A Mac that’s actually fast at inference	Mac Mini M5 Pro	from $1,699	Check price
Not sure the workload justifies any purchase yet	Rented GPU, ~$1/hr	pay per hour	RunPod

Skip the $899 and $1,099 configs for AI entirely: 16GB can’t hold the models worth running, and if you’re paying $1,099 for 24GB you’re $200 from the config that gets both the full 27B class and Ollama’s MLX default.

FAQ

Is the M6 Mac Mini faster than the M4 Pro Mini for local AI? No — and this trips people up. The M4 Pro’s 273 GB/s beats the M6’s 170 GB/s by 60%, so a clearance M4 Pro Mini decodes faster than the brand-new M6. The M6 wins on memory ceiling (32GB vs the M4 Pro’s common 24GB config), single-threaded CPU speed, and Neural Engine compute — none of which set your tok/s. If you find an M4 Pro 24GB discounted below $1,100, it’s the better pure-inference buy; the M6 32GB is the better model-fit buy.

Can the M6 Mini run Qwen3.8-27B? The 32GB config, yes — the 16.8GB Q4_K_M sits well inside the ~24GB default GPU budget with headroom for 32K-class context. The 16GB and 24GB configs, no and barely: 16GB machines offer ~12GB to the GPU, and 24GB machines fit the weights with thin context margins. Expect ~8–10 tok/s (bandwidth estimate; MLX toward the top of that range) against 41 tok/s measured on a used RTX 3090.

Does the dual Neural Engine speed up Ollama or LM Studio? No. llama.cpp, MLX, and everything built on them run inference on the GPU cores; the Neural Engine serves Core ML and Apple Intelligence. The M6’s doubled Neural Engine compute is real hardware that today’s local-LLM stacks simply don’t touch — the same TOPS-vs-bandwidth distinction we documented for NPUs generally.

Should I wait for measured benchmarks before ordering? If you’re deciding between Mac configs, no — bandwidth math has predicted every M-series chip within ~10%, and 170 GB/s won’t surprise anyone. If you’re deciding Mac vs GPU, also no, because the gap isn’t close: the 3090 wins speed by 4–5× and the Mini wins power by ~10× regardless of where the M6 lands within its estimate range. The only buyers who should wait are M5 Pro considerers curious whether its “4× faster LLM processing” claim (Apple’s number, prefill-flavored) holds up in third-party MLX runs after September 22.

Is 32GB enough, or should I stretch to the M5 Pro’s 64GB? Depends on the model class you live in. 32GB covers everything through the 35B-MoE tier — the practical sweet spot of late 2026. 64GB opens dense 70B at Q4 (~8–12 tok/s at 307 GB/s, extrapolated) and the 100B-class MoE quants. If you know you want 70B+, the M5 Pro at 64GB (or a used Mac Studio) is the honest minimum; don’t buy 32GB planning to grow into models it can’t hold — nothing is upgradeable later.

Recommended Gear

• Mac Mini M6 32GB — the only M6 config worth buying for local AI: fits the 27B class at $1,299
• Mac Mini M5 Pro — 307 GB/s and up to 64GB; the Mac to buy if inference speed is the point
• Used RTX 3090 — still 4–5× the speed of any Mini on every model both can run

Sources

• Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute — Apple Newsroom
• Apple Reveals M6 as First-Ever 2nm Chip — MacRumors
• M6 Mac mini arrives with a new record-high starting price — AppleInsider
• New Mac mini M6 and M5 Pro: Everything you need to know — Macworld
• Mac mini roundup (M6 / M5 Pro configs and pricing) — MacRumors
• 2027 Macs to Get AI-Focused M7 Chips as Apple Skips High-End M6 — MacRumors
• Mac mini power consumption and thermal output — Apple Support
• Performance of llama.cpp on Apple Silicon M-series — llama.cpp discussions
• iogpu.wired_limit_mb on Mac: Raising the Metal Memory Ceiling — ModelPiper
• New and used RTX 3090 price history — BestValueGPU
• Used RTX 3090 price index — ResalePrices

Last updated September 8, 2026. Prices and specs change; verify current rates before purchasing.