Mac Studio M5 Ultra for local AI: 200B-plus models on one quiet desktop (2026)

By /
Mac Studio M5 Max and M5 Ultra prices on Apple UK, October 2026

The Mac Studio with M5 Ultra is Apple's most powerful desktop, with 96 GB or 256 GB of unified memory today and 512 GB due in late October. For local AI, buy the 256 GB version. It holds mixture-of-experts models of 200-300B, such as DeepSeek V4 Flash, and runs Llama 3.3 70B at about 23 tokens a second. Expect £9,499 at Apple UK. Our rating is 4/5. Still comparing? Start with 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 Studio M5 Ultra at a glance for local AI: £5,499 with 96 GB, up to 512 GB of memory, runs 235B mixture-of-experts models on 256 GB, rated 4 out of 5
Mac Studio M5 Ultra at a glance. Prices from UK stores, checked 17 October 2026.

Key facts

  • What it is: Apple's top desktop Mac, in the same 19.7 cm square case as the M5 Max model but heavier at 3.6 kg. It went on sale on 22 September 2026, and the CPU and GPU share one pool of unified memory.
  • Chip: Apple M5 Ultra with a 30-core CPU and 64-core GPU, or a 36-core CPU and 80-core GPU, both with Neural Accelerators in each GPU core and a 32-core Neural Engine.
  • Memory options: 96 GB or 256 GB on either chip; 512 GB only with the 36-core CPU and 80-core GPU, coming late October 2026 and not yet on sale in the UK store. Memory cannot be added after purchase.
  • Usable for AI: about 192 GB of the 256 GB can go to the GPU by default, because macOS caps the GPU at about three quarters of memory. One Terminal command raises it to about 240 GB.
  • Memory bandwidth: 1.2 TB/s (about 1,200 GB/s) on both chips.
  • UK price: from £5,499 (96 GB, 1 TB SSD); £9,499 for 256 GB with a 1 TB SSD, as checked 17 October 2026. No UK price yet for 512 GB.
  • Where to buy: Apple UK for every version; Scan takes pre-orders for the 96 GB base model.
  • Best for: running 200-300B mixture-of-experts models, and 70B dense models at a comfortable speed, on one machine.
  • Our rating (desk review, not hands-on): ★★★★☆ 4/5

On this page

Which AI models can the Mac Studio M5 Ultra run?

The 256 GB version holds models up to about 300B parameters at 4-bit, which covers Qwen3 235B-A22B, DeepSeek V4 Flash and MiniMax M2.5.

Memory optionBiggest model at 4-bitExample modelsPublished speed
96 GBAbout 115B (about 72 GB usable)Llama 3.3 70B (42.4 GB), gpt-oss-120b, Qwen3.5-122B-A10B after raising the limitgpt-oss-120b, MXFP4, oMLX: 90.2 tokens a second at 8K context, 64-core GPU (oMLX community)
256 GBAbout 315B (about 192 GB usable), about 395B after raising the limitQwen3 235B-A22B (142.5 GB), DeepSeek V4 Flash (284B), MiniMax M2.5 (229B), Llama 3.3 70B at 8-bitLlama 3.3 70B, 4-bit, oMLX: 23.5 tokens a second at 4K (oMLX community); DeepSeek V4 Flash, MXFP4, oMLX: 61.7 at 8K (oMLX community)
512 GBAbout 635B (about 384 GB usable), about 820B after raising the limit671B models such as DeepSeek V3.1 at 4-bit, after raising the limitNo published figure; not shipping until late October

Long conversations need 2-4 GB more and slow generation: Llama 3.3 70B fell from 23.5 tokens a second at 4K context to 18.7 at 32K in the same oMLX session. Mixture-of-experts models read only their active parameters for each token, which is why DeepSeek V4 Flash (284B total, 13B active) writes faster than the much smaller dense 70B; 1-trillion-parameter models such as Kimi K2 Thinking (about 594 GB) need two 512 GB machines.

How fast is it?

Token generation on a Mac is limited mostly by memory bandwidth, Apple's machine learning team says, and the M5 Ultra has twice the M5 Max's 614 GB/s. The new GPU Neural Accelerators mainly speed up reading the prompt. Published figures for the 256 GB M5 Ultra:

  • Llama 3.3 70B, 4-bit, oMLX 0.7.0rc1 (MLX), macOS 27.0.1, 64-core GPU: 24.5 tokens a second at 1K context, 23.5 at 4K and 18.7 at 32K; it read prompts at 569 tokens a second at 4K, with 39 GB peak memory. oMLX community benchmark.
  • DeepSeek V4 Flash (284B total, 13B active), MXFP4 4-bit, oMLX 0.7.0, 64-core GPU, multi-token prediction on: 61.7 tokens a second at 8K context and 49.2 at 128K, with 158 GB peak memory. oMLX community benchmark.
  • MiniMax M2.5 (229B), 4-bit, oMLX 0.7.0rc1, 64-core GPU: 28.3 tokens a second at a 64K-token context. oMLX community benchmark.
  • Qwen 3.5 122B A10B, 36-core CPU model: about 80 tokens a second, against about 60 on the M3 Ultra; BGR did not state the quantisation or tool. BGR review.
  • GLM-5.3-Flash, mixed 4-bit and 8-bit, oMLX, 80-core GPU: 31 tokens a second writing and 1,217 reading a 3,408-token prompt, against 21 and 458 on an M3 Ultra. MacStories found prompt reading about 150% faster than the M3 Ultra on average. MacStories review.
  • Llama 2 7B, Q4_0, llama.cpp (llama-bench), 80-core GPU: 179.10 tokens a second when writing and 4,944.66 when reading the prompt. llama.cpp Apple Silicon discussion.

Nobody has published figures for the 512 GB version yet. Against others on our list, the M5 Ultra is about 1.8 times as fast as the Mac Studio M5 Max on Llama 3.3 70B (23.5 against 12.6 tokens a second in oMLX). Tom's Hardware found its token throughput almost four times that of the NVIDIA DGX Spark, though its exact figures appear only in charts.

Which version should you buy?

Buy the 256 GB version with the 30-core CPU and 64-core GPU at £9,499. It has the same 1.2 TB/s bandwidth as the 80-core chip, and bandwidth is what sets writing speed. The 80-core version costs £1,300 more (£10,799) and mainly speeds up prompt reading, which matters if you feed it long documents or code.

Skip the 96 GB Ultra at £5,499: the M5 Max with 128 GB costs £400 less and holds more. Wait for 512 GB only if you need 400B-plus models at 4-bit on one machine; it requires the 36-core CPU and 80-core GPU, and Apple has not yet published its UK price.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
M5 Ultra, 30-core CPU, 64-core GPU96 GB1 TB£5,499Apple UK, Scan (pre-order)
M5 Ultra, 30-core CPU, 64-core GPU256 GB1 TB£9,499Apple UK
M5 Ultra, 30-core CPU, 64-core GPU256 GB2 TB£9,999Apple UK
M5 Ultra, 36-core CPU, 80-core GPU96 GB1 TB£6,799Apple UK
M5 Ultra, 36-core CPU, 80-core GPU256 GB1 TB£10,799Apple UK
M5 Ultra, 36-core CPU, 80-core GPU512 GB1 TBNot yet priced; due late OctoberApple UK

All prices include VAT. Apple's education price starts at £5,099 for the base M5 Ultra. Scan matches Apple's £5,499 for the base model; I found no UK discounts.

Hidden costs

  • Memory you can't add later: going from 96 GB to 256 GB costs £4,000 at purchase, and there is no upgrade afterwards.
  • Storage: over the 1 TB base, Apple charges £500 for 2 TB, £1,500 for 4 TB, £3,500 for 8 TB and £7,500 for 16 TB. DeepSeek V4 Flash at 4-bit used about 158 GB of memory in one published test, so a few models of that size fill 1 TB quickly.
  • AppleCare+: optional insurance on top of the one-year warranty. Each accidental damage claim carries an excess of £79 for enclosure damage or £229 for other damage.
  • Power: Apple measures 9 W idle and up to 385 W flat out for a 512 GB, 16 TB M5 Ultra. Budget for that if it will run long jobs every day.
  • Screen, keyboard and mouse: the box holds the Mac Studio and a power lead, nothing else. A cluster also needs a Thunderbolt 5 cable between every pair of Macs.

Software: Ollama, LM Studio and the rest

  • MLX: Apple's open-source machine learning framework for Apple silicon, which Apple promotes for running and fine-tuning models on the Mac Studio. From macOS 26.2 it uses the M5 GPU's Neural Accelerators.
  • LM Studio: runs both MLX and llama.cpp engines and is the easiest place to start. Apple shows LM Studio Bionic in its M5 Ultra marketing. See our LM Studio guide.
  • Ollama: moved to an MLX engine on Apple silicon with version 0.19 in March 2026, launched as a preview. See our Ollama guide.
  • llama.cpp: runs on the GPU through Metal, and GGUF files work as on any Apple silicon Mac.
  • oMLX: an open-source MLX server for Apple silicon; most published M5 Ultra speeds above come from its community benchmarks.
  • No CUDA: tools that need an NVIDIA GPU will not run.
  • One setup step for big models: to give a 256 GB Mac Studio about 240 GB of GPU memory, run sudo sysctl iogpu.wired_limit_mb=245760 in Terminal and leave 8-16 GB for macOS. The setting resets when you restart. The mlx-lm instructions recommend the same setting when a model is close to the limit.
  • Clustering: exo joins several Macs over Thunderbolt 5 with RDMA. It needs macOS 26.2 or later, RDMA switched on with rdma_ctl enable in Recovery, every Mac cabled to every other and the same macOS build on each. Apple says four M5 Ultra Mac Studios ran a 72B model up to 3x faster than one. Jeff Geerling's tests on four M3 Ultra Mac Studios reached about 32 tokens a second on Qwen3 235B and about 30 on Kimi K2 Thinking with exo, but he also hit failed model loads on prerelease software.

Noise, power and size

  • Size and weight: 19.7 x 19.7 x 9.5 cm and 3.6 kg.
  • Power supply: built in, with 480 W maximum continuous power.
  • Power draw: Apple lists 9 W idle and 385 W maximum for the 36-core, 80-core GPU model with 512 GB and 16 TB.
  • Noise and heat: MacStories could not hear the fan in daily use without putting an ear to the case. Tom's Hardware found it largely quiet, though the fan can be heard, with the CPU averaging 78.29°C in a 10-run Cinebench 2026 stress test.

Who makes the Mac Studio M5 Ultra?

  • Apple Inc., founded in 1976 and based in Cupertino, California.
  • In the UK it is sold by Apple online, in the Apple Store app and in Apple Stores, and by retailers such as Scan.
  • Apple handles support: a one-year limited warranty, on top of your rights under UK consumer law.

What's good

  • 256 GB of unified memory today and 512 GB from late October, the most of any machine on our list.
  • 1.2 TB/s of bandwidth: Llama 3.3 70B at 4-bit runs at about 23 tokens a second.
  • Mixture-of-experts models of 229-284B run at about 28-67 tokens a second on one desktop, depending on model and context length.
  • Prompt reading is much faster than the M3 Ultra, about 150% on average in MacStories' tests.
  • Six Thunderbolt 5 ports, which leaves room to cluster Macs over RDMA.
  • Sold in the UK by Apple with VAT included, and Apple takes change-of-mind returns within 14 days.

What's not so good

  • £9,499 for 256 GB, and the memory upgrade from 96 GB alone costs £4,000.
  • The 512 GB version is not on sale yet and has no UK price; Tom's Hardware reported a 16-18 week wait for the 256 GB model it tested.
  • No CUDA, and an RTX 5090 is still faster on models that fit in its 32 GB: MacStories measured 59 against 48 tokens a second on Qwen3.8-27B.
  • Storage upgrades are expensive, up to £7,500 for 16 TB.
  • The 96 GB base model has less memory than the top M5 Max.

Who should buy it?

The 256 GB M5 Ultra suits developers, researchers and small teams who want 200B-plus open models, or several large models at once, running privately on one quiet machine. Skip it if 70B or 120B models are enough: the Mac Studio M5 Max with 128 GB costs about half as much. If your tools need CUDA, look at the NVIDIA DGX Spark instead.

Buying in the UK

  • Availability: Apple UK sells every version except 512 GB, which is due in late October. Scan takes pre-orders for the 96 GB base model. The 256 GB version is built to order, and I found it only at Apple. Check Apple's delivery date before ordering.
  • VAT and import charges: UK prices include VAT and there is nothing more to pay. Importing a US machine rarely saves money: GOV.UK says VAT is charged on goods over £135 on the price plus postage, insurance and any duty, and the courier collects it before delivery.
  • Warranty: Apple gives a one-year limited warranty, separate from and in addition to UK consumer law. Scan lists a 12-month warranty with a 28-day dead-on-arrival period.
  • Returns: Apple accepts change-of-mind returns within 14 calendar days of delivery; custom-configured Macs may not be exchanged in an Apple Store.
  • 2026 price changes: Apple raised UK Mac prices on 25 June 2026 because of the memory shortage. Hoxton Macs reports the M3 Ultra Mac Studio went from £3,999 to £5,299 after its high-memory versions were withdrawn in May. The M5 Ultra starts £200 higher, at £5,499.

Interesting facts about the Mac Studio M5 Ultra

  • The 512 GB option needs the top chip with the 36-core CPU and 80-core GPU; the 30-core version stops at 256 GB.
  • Apple's "3x faster" cluster claim was measured with a 72B dense model, a 32K-token prompt and mlx-lm on four Mac Studios, each wired to the other three over Thunderbolt 5 with RDMA.
  • In Apple's own power table, a 512 GB M5 Ultra draws up to 385 W, about 40% more than the 270 W of the M3 Ultra with the same memory.

Alternatives to the Mac Studio M5 Ultra

  • Mac Studio M5 Max - pick it if 128 GB is enough; it runs 70B and 120B-class models for £5,099.
  • NVIDIA DGX Spark - pick it if your tools need NVIDIA's CUDA.
  • MacBook Pro M5 - pick it if you want the M5 Max chip with up to 128 GB in a laptop.

Our rating

AreaScore
Memory for AI5/5
Speed5/5
Value3/5
Software4/5
UK buying4/5
Overall★★★★☆ 4/5

The 256 GB Mac Studio M5 Ultra is the fastest single machine on our list for 200B-plus models. It is expensive, and most people running 70B or 120B models will be happy with the M5 Max.

Frequently asked questions

Can the Mac Studio M5 Ultra run DeepSeek?

Yes. With 256 GB it runs DeepSeek V4 Flash (284B) at 4-bit, at about 50-67 tokens a second with 1K-32K of context in published oMLX tests, with multi-token prediction switched on. The larger 671B DeepSeek V3.1-class models need the 512 GB version.

Can you buy the 512 GB Mac Studio M5 Ultra in the UK?

Not yet. Apple lists 512 GB in the specifications and says it is coming in late October 2026, but on 17 October the UK store offered only 96 GB and 256 GB, and there was no UK price.

Is the Mac Studio M5 Ultra better than the M5 Max for AI?

It is about 1.8 times as fast on 70B models and holds twice the memory or more. If your models fit in 128 GB, the M5 Max gives most of the benefit for about half the price.

Sources

More guides

Scroll to Top