Which Mac mini Should You Buy for Local LLMs? (2026)

Bottom Line

• Best Mac mini for local LLMs: M6, 32GB, 256GB, $1,299. It has the lowest price per GB in the lineup (derived: $40.6). It runs a 27B model at Q4 with context. Model picks per tier are on the Mac mini local LLM hub.
• Speed pick: M5 Pro, 15-core CPU / 16-core GPU, 48GB, $2,299. Apple lists 307 GB/s, 1.8x the M6’s 170 GB/s (derived: 307 ÷ 170).
• Largest memory: the same M5 Pro with 64GB, $2,699. It still does not run gpt-oss 120B. Those weights are 65.25GB, more than the machine’s 64GB.
• Never order the $899 16GB M6 for local AI. Apple lists it at 153 GB/s, not 170. The $200 step to 24GB adds 8GB and the faster bus.
• Skip the $200 20-core GPU step on the M5 Pro. Apple lists one bandwidth, 307 GB/s, for the M5 Pro. The step buys prompt processing, not token speed.
• Every memory step costs $25 per GB. The rate is the same on both chips. The only big price jump is the chip itself.

Prices are Apple US prices, as of September 2026, read from Apple’s US configurator on 22 September 2026. Apple lists machines in stores from 22 September 2026. The M4 Mac mini is discontinued.

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

The Decision Nobody Frames Correctly

Most Mac mini pages compare the M6 and the M5 Pro as two products. The configurator sells nine configurations that matter for local AI. Three numbers separate them: memory, bandwidth and price per GB.

Bandwidth comes in three fixed steps: 153 GB/s, 170 GB/s and 307 GB/s. Memory comes in five sizes: 16GB, 24GB, 32GB, 48GB and 64GB. The M6 stops at 32GB. The M5 Pro starts at 24GB.

Apple prices memory at a flat $25 per GB on both chips. So the real decision is one step: from the 32GB M6 at $1,299 to the 48GB M5 Pro at $2,299. That step costs $1,000. At Apple’s own rates, $400 of it pays for 16GB of memory and $200 pays for the 256GB-to-512GB storage step. The remaining $400 buys the M5 Pro chip and 1.8x the bandwidth (derived).

The model picks for each tier are on the Mac mini local LLM hub. This page only tells you which line to click.

Every Configuration, Priced

All prices are US, base storage, as of September 2026. Price per GB is a derived figure: price divided by unified memory in GB. Usable memory assumes macOS gives the GPU roughly 75% of unified memory, a community-reported approximation.

Bandwidth comes from Apple’s Mac mini specs page. It lists 153 GB/s for the 16GB M6, 170 GB/s for the 24GB and 32GB M6, and 307 GB/s for the M5 Pro. It gives no separate figure for the 20-core GPU option.

The arithmetic for the derived column: $899 ÷ 16 = $56.2, $1,099 ÷ 24 = $45.8, $1,299 ÷ 32 = $40.6, $1,699 ÷ 24 = $70.8, $2,299 ÷ 48 = $47.9, $2,699 ÷ 64 = $42.2, $1,899 ÷ 24 = $79.1, $2,499 ÷ 48 = $52.1, $2,899 ÷ 64 = $45.3.

Two things stand out. The 32GB M6 is the cheapest memory in the line per GB. The 24GB M5 Pro is the most expensive memory per GB, and it holds the same models as the $1,099 M6.

What Fits Each Tier

File sizes come from the Hugging Face file listings, read on 22 September 2026. Usable memory is 75% of the installed memory: about 12.9GB, 19.3GB, 25.8GB, 38.7GB and 51.5GB (derived, 1 GiB = 1.074GB). “Tight” means less than 3GB is left for context.

The Qwen 3.8 27B row adds the 16.46GB model file and its 0.93GB vision encoder. The gpt-oss 20B row uses OpenAI’s own checkpoint; the Q4_K_M GGUF is 11.62GB and fills a 16GB mini on its own.

The Four Traps in the Configurator

Trap 1: the 16GB M6 has a slower bus

The $899 base M6 is the only Mac mini at 153 GB/s. The 24GB and 32GB M6 run at 170 GB/s. So the $200 step from 16GB to 24GB buys 8GB of memory and an 11% faster bus (derived: 170 ÷ 153 = 1.11). The 16GB M6 is a 9B-class machine with about 12GB usable. Do not buy it for local AI.

Trap 2: the 20-core GPU M5 Pro

The step to the 18-core CPU / 20-core GPU costs $200 at every memory size. Derived: $1,899 minus $1,699, $2,499 minus $2,299, $2,899 minus $2,699. Apple lists 307 GB/s for the M5 Pro and no higher figure for this option. Token generation is bandwidth-bound, so the extra GPU cores do not raise tokens per second. They speed up prompt processing, which is compute-bound. Put the $200 toward memory instead: it buys 8GB at Apple’s rate.

We also checked for a Mac Studio-style bin lock. There is none. The cheaper 15-core M5 Pro takes 48GB and 64GB, at $2,299 and $2,699.

Trap 3: 64GB does not unlock gpt-oss 120B

This is the question most 64GB buyers ask. The answer is no. OpenAI’s gpt-oss 120B checkpoint is 65.25GB, which is 60.8 GiB. A 64GB Mac mini has 64 GiB in total. At the default ~75% GPU share, about 48 GiB is available, 12.8 GiB short.

You can raise the limit with sudo sysctl iogpu.wired_limit_mb=<value>. Community guides advise leaving 8GB to 16GB for macOS. At 56 GiB for the GPU, the model is still 4.8 GiB short, before any context. The Q4_K_M GGUF does not rescue it; it is 62.77GB. The cheapest new Apple box that runs gpt-oss 120B is the Mac Studio M5 Max 128GB at $5,099.

Trap 4: Apple storage

Apple’s storage steps, read on 22 September 2026: the M6 256GB-to-512GB step is $200. The 512GB-to-1TB step is $300 on both the M6 and the M5 Pro. A 27B model at Q4_K_M is 16.82GB, so 256GB holds several models. A Thunderbolt NVMe enclosure holds a larger library. Buy the base storage and spend on memory.

One Pick Per Budget

1. Around $1,100: M6, 24GB, $1,099. The floor. It runs gpt-oss 20B and a 27B at Q4 with little context left. Skip the 16GB base.
2. Around $1,300, the pick for most buyers: M6, 32GB, $1,299. The lowest price per GB in the line. It runs Qwen 3.6 27B at Q4_K_M with about 9GB left for context, or Qwen3.6-35B-A3B at Q4 with about 3.6GB left.
3. Around $2,300, speed first: M5 Pro, 15-core CPU / 16-core GPU, 48GB, $2,299. 1.8x the M6 bandwidth. It runs a 27B at Q8_0 (28.60GB) and Qwen3.6-35B-A3B at Q4 with long context.
4. Around $2,700, the ceiling: M5 Pro, 15-core CPU / 16-core GPU, 64GB, $2,699. It runs Qwen3.6-35B-A3B at Q8_0 (36.90GB) with room left. It does not run gpt-oss 120B.
5. The 24GB M5 Pro at $1,699: no. Same memory as the $1,099 M6. For $600 more ($25 per GB), the 48GB M5 Pro doubles it.
6. The 20-core GPU at any memory size: no. Same listed 307 GB/s. Save the $200.

If you need more than 64GB, the Mac mini is the wrong machine. The Mac Studio config picker covers the next step. The Mac Studio M5 Max with the 40-core GPU and 64GB is $3,499. For $800 more than the 64GB mini, it doubles the bandwidth to 614 GB/s.

The Discontinued Option

Apple replaced the M4 Mac mini on 25 August 2026. Apple’s support specs page lists the M4 at 120 GB/s with a 10-core CPU and 10-core GPU. The listings we found on 22 September 2026 were 16GB units, so treat it as a 12GB-usable, 9B-class machine.

Use one number to judge a listing. The M4 matches the $899 M6 on bandwidth per dollar at about $705 (derived: 120 ÷ 153 × $899). Below that, the M4 is the better local-LLM dollar. Above it, buy the new M6.

What we found on 22 September 2026:

• Costco: $479.99 for members, in-warehouse only, per the Warehouse Runner stock tracker. The tracker showed stock at 10 of 610 warehouses.
• Amazon: the listing showed one used unit from a third-party seller, priced above the $705 line.

Stock is nearly gone. If you find a new unit under $705, the Mac mini M4 is still worth it. Check the condition, seller and memory line before you order. Otherwise buy the 24GB or 32GB M6.

How to Predict Your Own Speed

Rough ceiling in tokens per second = memory bandwidth ÷ size of the model file. A dense model reads every weight for each token, so this is the upper bound. Real speed is lower, because attention and the KV cache also cost time.

These are estimates from the formula, not bench runs. Mixture-of-experts models break the formula in your favour. Qwen3.6-35B-A3B reads only its active experts per token, so it runs much faster than a dense 27B of similar file size.

How to Read Your Own Order Page

Before you click Buy, check three lines on Apple’s configurator:

1. Memory on the M6. If it says 16GB, the bus is 153 GB/s and you have about 12GB for models. You want 24GB or 32GB.
2. Chip on the M5 Pro. If it says 18-core CPU / 20-core GPU, you paid $200 for prompt processing. The 15-core option has the same listed 307 GB/s.
3. Storage. If it is above the base tier, you paid $200 or $300 per step for space an external drive provides.

Memory is soldered. Order the tier for the largest model you intend to run in three years, not the one you run today.

See Also

• Best Local LLM for Mac mini — the hub: model picks for every memory tier
• Which Mac Studio Should You Buy for Local LLMs? — the same config picker, one tier up
• Mac mini vs GPU for Local LLM — the same money spent on a graphics card
• Edge0 35B-A3B on a 16GB Mac mini — whether SSD expert streaming rescues a 16GB mini
• Mac mini vs Mac Studio for Local LLMs — whether you need the Studio at all
• gpt-oss 120B vs 20B: Which to Run — why the 120B needs more than a Mac mini
• MoE vs Dense on 24GB — why 35B-A3B beats a dense 27B on speed