Mac mini M5 Pro for local AI: 70B models on a small desk (2026)

By /
Mac mini M5 Pro and M6 prices on Apple UK, October 2026

The Mac mini with M5 Pro is Apple's small desktop with nearly twice the M6 model's memory speed and up to 64 GB. Buy the 64 GB version: £2,699 from Apple with the 15-core CPU and 512 GB of storage. It holds Llama 3.3 70B at 4-bit and runs 32B models with long chats. The 24 GB model starts at £1,699 but holds less than a 32 GB M6. I rate it 4 out of 5. Compare it in which Mac or mini PC to buy for local AI.

Last updated 22 October 2026. Written by Annie, a search and digital marketing specialist. Based on the maker's specifications, UK prices and published benchmarks; I have not tested this machine myself. See the editorial policy.

Mac mini M5 Pro at a glance for local AI: £1,699 with 24 GB, up to 64 GB of memory, runs 70B models on 64 GB, rated 4 out of 5
Mac mini M5 Pro at a glance. Prices from UK stores, checked 17 October 2026.

Waiting for a deal? Black Friday is 27 November 2026. Verified UK price drops on the Mac mini, and the configurations worth holding out for, are on the Black Friday AI deals page.

Key facts

  • What it is: Apple's small desktop, 12.7 cm square, with the M5 Pro chip. Announced 25 August 2026 and on sale from 22 September 2026, replacing the M4 Pro model.
  • Chip: Apple M5 Pro, either a 15-core CPU (5 super cores, 10 performance cores) with a 16-core GPU, or an 18-core CPU (6 super cores, 12 performance cores) with a 20-core GPU. Each GPU core has a Neural Accelerator, and there is a 16-core Neural Engine.
  • Memory options: 24 GB, 48 GB or 64 GB of unified memory, and Apple sells 48 GB and 64 GB with either chip. You choose at purchase. It cannot be upgraded later.
  • Usable for AI: about 16 GB (24 GB model), 36 GB (48 GB) or 48 GB (64 GB) can go to the GPU by default: two thirds of the memory at 24 GB, three quarters above 32 GB. You can raise it with one Terminal command (see Software below).
  • Memory bandwidth: 307 GB/s
  • UK price: from £1,699 (15-core CPU, 24 GB, 512 GB), or £2,699 for 64 GB (15-core CPU, 512 GB, built to order), as checked 17 October 2026.
  • Where to buy: Apple UK (the only place to get 48 GB or 64 GB); Currys for the 24 GB model.
  • Best for: 32B models with long chats, Llama 3.3 70B at 4-bit, and a home AI server for agents and coding tools.
  • Our rating (desk review, not hands-on): ★★★★☆ 4/5

On this page

Which AI models can the Mac mini M5 Pro run?

macOS gives the GPU about two thirds of the memory on the 24 GB model and about three quarters on the 48 GB and 64 GB models, so plan around those figures.

Memory optionBiggest model at 4-bitExample modelsPublished speed
24 GB (about 16 GB for the GPU)14B with long chats (Qwen3 14B is 8.5 GB)Llama 3.1 8B (4.8 GB), Qwen3 14BLlama 2 7B, 4-bit (Q4_0), llama.cpp: about 68 tokens a second (16-core GPU) and 66 (20-core GPU) (llama.cpp Apple silicon thread)
48 GB (about 36 GB for the GPU)32B (19.4 GB) with long chatsQwen3.8 27B, 32B models, Qwen3 30B-A3B at 8-bit (about 32 GB)No published figure for these models on the M5 Pro
64 GB (about 48 GB for the GPU)70B (Llama 3.3 70B is 42.4 GB) with a normal-length chatLlama 3.3 70B, 32B models at 8-bit (about 34 GB)No published 70B figure for the M5 Pro

A long conversation adds 2-4 GB on top of the model file, which is why 70B is a tight fit even at 64 GB. Mixture-of-experts models such as Qwen3 30B-A3B read only about 3B parameters for each word, so they write far faster than a dense 32B model, but they still need memory for all their weights. The 235B Qwen3 model (142.5 GB at 4-bit) is out of reach.

How fast is it?

The best documented figures come from the llama.cpp project's Apple silicon benchmark thread, which tests every chip with the same small model.

  • Llama 2 7B, 4-bit (Q4_0), llama.cpp: about 68 tokens a second when writing on the 16-core GPU and 66 on the 20-core GPU. Prompt reading was about 1,340 and 1,621 tokens a second.
  • Llama 2 7B, 8-bit (Q8_0), llama.cpp: about 39 tokens a second on both versions.
  • Against the M4 Pro: the 20-core M4 Pro wrote about 51 tokens a second and read prompts at about 440 in the same thread. So the M5 Pro writes about 30% faster and reads prompts about 3.7 times faster.
  • Against other machines here: the Mac mini M6 wrote about 36 tokens a second (an M6 with an 8-core GPU), and the 40-core M5 Max in the MacBook Pro about 120.
  • Apple's claim: up to 4 times faster prompt processing in LM Studio than the M4 Pro Mac mini. Apple did not name the model.
  • 70B models: no published figure for the M5 Pro by 17 October 2026. Tech site ifanr measured about 10 tokens a second with Llama 3.3 70B at 4-bit on a 128 GB M5 Max in LM Studio. The M5 Pro has half that chip's memory bandwidth, so expect it to be clearly slower.

Writing speed tracks memory bandwidth because the chip reads the whole model for every word. Prompt reading tracks GPU power, which is where the extra £200 for the 20-core GPU shows.

Which version should you buy?

Buy the 64 GB model with the 15-core CPU and 16-core GPU, at £2,699 with 512 GB of storage. In the llama.cpp test both GPU versions wrote text at the same speed, because they share the same 307 GB/s memory. The 20-core GPU (£200 more) reads long prompts about a fifth faster, which helps if you paste in long documents or run coding agents over big projects.

If 32B models are your limit, the 48 GB model at £2,299 saves £400. Skip the 24 GB model for AI: it gives the GPU less memory than a £1,299 Mac mini M6 with 32 GB.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
15-core CPU, 16-core GPU, standard24 GB512 GB£1,699Apple UK, Currys
18-core CPU, 20-core GPU, built to order24 GB512 GB£1,899Apple UK
15-core CPU, 16-core GPU, built to order48 GB512 GB£2,299Apple UK
18-core CPU, 20-core GPU, built to order48 GB512 GB£2,499Apple UK
15-core CPU, 16-core GPU, built to order (our pick)64 GB512 GB£2,699Apple UK
18-core CPU, 20-core GPU, built to order64 GB512 GB£2,899Apple UK
18-core CPU, 20-core GPU, standard64 GB1 TB£3,199Apple UK

Students and staff pay from £1,599, and £2,499 for our 64 GB pick, in Apple's UK education store. Apple's UK refurbished store listed no Mac mini on 17 October 2026.

Hidden costs

  • Memory you can't add later. Going from 24 GB to 48 GB costs £600 and to 64 GB costs £1,000, all at purchase.
  • Storage. 512 GB is the base. Apple charges £300 more for 1 TB, and large model files fill a drive quickly.
  • Screen, keyboard and mouse. None are included. The box holds the Mac mini and a power lead.
  • AppleCare. Optional. AppleCare One costs from £16.99 a month for up to three products. Accidental damage claims carry an excess of £39-£79 for casing damage and £109-£229 for other damage such as liquid.
  • Clustering. Apple's idea of linking Mac minis over Thunderbolt 5 for bigger models means buying a second machine at £1,699 or more.

Software: Ollama, LM Studio and the rest

Everything runs natively on Apple silicon, with no drivers to install.

  • MLX: Apple's own machine learning framework runs on every Apple silicon Mac. Apple says MLX uses the M5 family's GPU Neural Accelerators from macOS 26.2.
  • Ollama: runs models through llama.cpp with Metal. Its MLX engine, first released as a March 2026 preview, asked for more than 32 GB of memory, which the 48 GB and 64 GB models meet. See our Ollama guide.
  • LM Studio: uses both MLX and llama.cpp and needs macOS 14 or later. See our LM Studio guide.
  • Clustering: Apple says Thunderbolt 5 lets you link several M5 Pro Mac minis to run large models on your own machines. Jeff Geerling's tests of this idea used the free Exo software, with RDMA over Thunderbolt 5 turned on from Recovery mode in macOS 26.2.
  • One setup step: to give a 70B model more room on the 64 GB model, run sudo sysctl iogpu.wired_limit_mb=53248 in Terminal (that sets 52 GB). It resets when you restart unless you add it to /etc/sysctl.conf. Leave several gigabytes for macOS.

Noise, power and size

  • Size: 12.7 x 12.7 x 5.0 cm and 0.73 kg. The power supply is built in.
  • Power and noise: Apple quotes maximum continuous power of 155 W and 5 dBA at idle for the Mac mini.
  • Under load: Macworld's review found the fan gets loud under demanding games and described the small case's cooling as limited.

Who makes the Mac mini M5 Pro?

  • Apple Inc., based at One Apple Park Way, Cupertino, California. Apple introduced the Macintosh in 1984.
  • UK support comes through Apple Stores, the online store and 0800 048 0408.
  • Apple gives a one-year limited warranty. Under English law you also have up to six years to claim from the seller for faulty goods.

What's good

  • 64 GB holds Llama 3.3 70B at 4-bit, and 48 GB holds 32B models with long chats.
  • 307 GB/s memory: about 67 tokens a second on the llama.cpp 7B test, nearly double the M6.
  • Prompt reading about 3.7 times the M4 Pro's in that test, which helps with long documents.
  • Three Thunderbolt 5 ports, with clustering for larger models backed by Apple.
  • Ollama, LM Studio, MLX and llama.cpp all run natively, in a 0.73 kg box with a built-in power supply.

What's not so good

  • 48 GB and 64 GB are built to order, so only Apple sells them, and the jump from 24 GB to 64 GB costs £1,000.
  • No published 70B speed yet, and half the M5 Max's bandwidth means 70B answers will be slow.
  • Macworld called it a good computer at a bad price, after its US review model cost $600 more than the M4 Pro equivalent.
  • The fan gets loud under heavy gaming loads, according to Macworld.
  • 512 GB of base storage.

Who should buy it?

Buy it if you want 32B to 70B models at home on a desk, for example a private research assistant or a local coding agent that runs all day. Skip it if you only need 8B-27B models: the Mac mini M6 with 32 GB does that for £1,299. If you want 128 GB, look at the MacBook Pro with M5 Max.

Buying in the UK

  • Availability: Apple UK sells every version. Currys stocks the 24 GB model at £1,699. The 48 GB and 64 GB versions are built to order by Apple. Prices include 20% VAT and there are no import charges.
  • Warranty: Apple gives one year; Currys lists the 1-year manufacturer's guarantee.
  • Returns: Apple accepts returns within 14 days of delivery.
  • Education: Apple's UK education store takes £200 off our 64 GB pick for students and staff.

Interesting facts about the Mac mini M5 Pro

  • In the llama.cpp test the 16-core and 20-core GPU versions wrote at the same speed (about 68 and 66 tokens a second), because their memory is identical. Only prompt reading improved.
  • Jeff Geerling used RDMA over Thunderbolt 5 to link four Mac Studios with M3 Ultra chips, and ran Qwen3 235B at about 32 tokens a second with Exo. Apple now pitches the same Thunderbolt 5 link for M5 Pro Mac minis.
  • The model it replaced, the M4 Pro Mac mini, rose from £1,399 to £1,599 in Apple's June 2026 UK price rise, according to Hoxton Macs. The M5 Pro model launched at £1,699.

Alternatives to the Mac mini M5 Pro

Our rating

AreaScore
Memory for AI3/5
Speed3/5
Value3/5
Software5/5
UK buying5/5
Overall★★★★☆ 4/5

With 64 GB the Mac mini M5 Pro is a small, quiet way to hold a 70B model at home, and every Mac AI tool works on it. You pay Apple's high price for memory, and 70B answers will not be quick.

Frequently asked questions

Can the Mac mini M5 Pro run Llama 70B?

Yes, with 64 GB. Llama 3.3 70B at 4-bit is 42.4 GB, under the roughly 48 GB the GPU gets by default. Nobody had published a speed for it on the M5 Pro by 17 October 2026, and 48 GB is not enough.

Should I buy the 48 GB or 64 GB Mac mini M5 Pro?

Buy 64 GB if you want 70B models. Buy 48 GB, which is £400 cheaper, if 32B models with long chats are your limit.

Is the Mac mini M5 Pro worth it over the M6 for AI?

Yes, if you want models bigger than 27B or faster answers. It has 307 GB/s memory against 170 GB/s, and wrote about 67 tokens a second against about 36 on the same llama.cpp test.

Sources

More guides

Scroll to Top