Best Local LLM for Mac mini (2026): Why 24GB Is the Floor

Bottom Line (September 2026)

• 16GB is not 16GB. macOS hands the GPU about 75% of unified memory by default. A 16GB Mac mini gives you roughly 12GB for models.
• Best pick on a 16GB mini: Qwen 3.5 9B at Q8_0, about 10GB. Near-full-precision quality with room left for context.
• 24GB is the floor and 32GB is the answer. 32GB gives you about 24GB usable, which is the first tier that runs a 27B model comfortably.
• The 2026 refresh undid both of Apple’s May 2026 memory cuts. The M6 goes to 32GB again. The M5 Pro goes to 64GB again.
• Entry price moved $599 → $799 → $899. But the M6 gives up to 170 GB/s against the M4’s 120 GB/s, so bandwidth per dollar improved at the top of the base model.
• The best local-LLM dollar is the discontinued M4, near $499 at retail. You lose about 30% of the bandwidth and save about 45% of the price.

Ready to buy? See the tested hardware list with current prices.

Prices are US street prices, as of September 2026. The memory market moves monthly.

The Lineup Changed on 25 August 2026

Most pages answering “best local LLM for Mac mini” still describe a machine Apple stopped selling. Here is the current lineup.

Apple announced the refresh on 25 August 2026, with machines in stores on 22 September 2026.

Two of those numbers matter more than the rest. In early May 2026 Apple cut the 64GB option from the M4 Pro Mac mini and cut the base M4 mini down to 16GB or 24GB. For four months, a 64GB Mac mini did not exist at any price. The 2026 refresh reverses both cuts. If you postponed a purchase because the ceiling was too low, the ceiling moved.

The Number Nobody States

You do not get all of your unified memory. macOS reserves roughly 25% of it for the system and gives the GPU about 75%. On a 16GB Mac mini that leaves about 12GB, and that single fact invalidates most 16GB model advice written for PCs.

You can raise the allocation. Setting iogpu.wired_limit_mb to about 13000 on a 16GB machine recovers roughly another gigabyte. It helps, and it does not change the tier you are in.

Picks by Memory Tier

16GB M6 ($899) — about 12GB usable

That last row is the trap. gpt-oss 20B is the correct pick on a 16GB PC and the wrong one on a 16GB Mac, because the PC gives the model close to its full 16GB and the Mac does not. The model is a mixture-of-experts design with 20.9B total parameters and only 3.6B active per token, which is why it is recommended so widely for 16GB machines — but the recommendation assumes 16GB is available.

24GB and 32GB M6 — the tier worth paying for

32GB is the configuration to buy. About 24GB of usable memory runs Qwen 3.6 27B at Q4 (~16GB) with a real context window still available, and it clears Qwen 3.8 27B at Q4_K_M (~18GB, including a 931MB vision encoder) with room to spare.

At 24GB you have about 18GB usable. Qwen 3.8 27B technically fits and leaves you very little. This is why we call 24GB the floor rather than the recommendation.

Apple prices these upgrades in its own configurator, and the step costs move. Check the current figure before you commit — a memory upgrade on Apple silicon is soldered and permanent.

48GB and 64GB M5 Pro ($1,699+) — 307 GB/s

The M5 Pro is a different class of machine: 307 GB/s against the M6’s 170 GB/s, roughly 1.8x. Bandwidth is what sets token speed, so a 27B model that runs on both machines runs almost twice as fast here.

At 64GB (about 48GB usable) you can hold a 35B-class model with a long context, or run a 27B model and a separate embedding or vision model at the same time. A dense 70B at Q4 is about 40GB of weights, which fits and then leaves almost nothing for context — the same trap the 48GB M4 Pro had, moved up one tier.

How to Predict Your Own Token Speed

Use this and you will not need anyone’s benchmark:

Rough ceiling in tokens/sec = memory bandwidth ÷ size of the model file.

The machine must read every active weight for each token it writes, so bandwidth divided by weight size is the hard upper bound. Expect 60% to 70% of that number in practice, because attention, the KV cache, and sampling all cost time.

Worked example on the M5 Pro at 307 GB/s with Qwen 3.6 27B at Q4 (~16GB): 307 ÷ 16 ≈ 19 tok/s ceiling, so 12–14 tok/s realistic. That tracks the measured 15–18 tok/s we have recorded for the same model on 273 GB/s hardware, which is the check that tells you the formula is sound.

On the 16GB M6 at 170 GB/s with Qwen 3.5 9B Q8_0 (~10GB): 170 ÷ 10 = 17 tok/s ceiling, so 10–12 tok/s realistic. That is comfortable reading speed and it is honest.

These are derived estimates, not bench runs, and they are marked as such.

The Honest Recommendation

If you want the most local AI per dollar right now, buy the outgoing M4 while retailers still have it. It has been selling near $499, which is below its own $599 launch price and well under the $799 it had climbed to before replacement. Retailers still list it — check the current Mac mini M4 price before you pay $899 for the M6. You accept 120 GB/s instead of 170 GB/s — roughly 30% less speed for roughly 45% less money — and a hard 24GB memory cap.

Two conditions make that the wrong choice. Buy new if you need 32GB or more, because the M4 cannot reach it. Buy new if this machine has to last, because clearance stock has no upgrade path and no long support runway.

If you are buying new, buy the 32GB M6, not the 16GB M6. The $899 entry model is a 9B machine. The step to 32GB is what turns the Mac mini into a local-LLM box, and it is the only memory decision you get to make — the memory is soldered.

If you need 64GB, the M5 Pro at $1,699 is now the answer, and four months ago there was no answer. Compare it honestly against a 128GB unified-memory box before you commit, because that tier gives you roughly twice the memory for a similar outlay.

See Also

• Mac mini M4 Pro: 48GB vs 64GB — the tier this refresh reopened
• Mac mini vs GPU for Local LLM — the same money spent on a graphics card instead
• Mac mini vs Mac Studio for Local LLMs — the next step up in the Apple line
• Best Local LLM for a 16GB MacBook Pro — the same 75% allocation problem, on a laptop
• Edge0 35B-A3B on a 16GB Mac mini — whether SSD expert streaming replaces buying RAM
• Best Local LLM for Mac Studio M4 — when the mini is not enough
• Should You Wait for the M5 Ultra Mac Studio? — the timing question one tier up
• What Local LLM Fits My Machine? — the general sizing tool
• OpenClaw on a Mac mini: Setup — what to do after the box arrives