Minisforum MS-S1 MAX for local AI: 128 GB from UK stock, or 192 GB (2026)

Minisforum MS-S1 MAX product page on Minisforum UK, October 2026

The Minisforum MS-S1 MAX is a compact AMD workstation with dual 10 Gb Ethernet and a PCIe slot. Most buyers should get the 128 GB model at £3,519 from Minisforum's UK warehouse: up to 96 GB goes to the GPU, enough for 70B models at 8-bit. The MS-S1 MAX-P495 has 192 GB, with up to 160 GB for the GPU, for £6,679 shipped from Hong Kong. We rate it 3.5/5. See our guide to which Mac or mini PC to buy for local AI.

Minisforum MS-S1 MAX at a glance for local AI: £2,359 with 64 GB, up to 192 GB of memory, runs 70B models on 128 GB, rated 3.5 out of 5
Minisforum MS-S1 MAX at a glance. Prices from UK stores, checked 17 October 2026.

Key facts

  • What it is: A small workstation with a built-in 320 W power supply that stands on a desk or goes in a 2U rack, sold direct by Minisforum UK.
  • Chip: AMD Ryzen AI Max+ 395 with Radeon 8060S graphics (40 compute units) on the 64 GB and 128 GB models. The MS-S1 MAX-P495 uses the newer Ryzen AI Max+ PRO 495 with Radeon 8065S graphics (40 compute units, up to 3.0 GHz).
  • Memory options: 64 GB or 128 GB of LPDDR5X-8000, or 192 GB of LPDDR5X-8533 on the P495. All are soldered, so you cannot add more later.
  • Usable for AI: up to 96 GB of the 128 GB can be set as graphics memory in the UEFI (BIOS), and 48 GB of the 64 GB. Minisforum says up to 160 GB of the P495's 192 GB can be allocated as graphics memory.
  • Memory bandwidth: 256 GB/s (Ryzen AI Max+ 395); about 273 GB/s (PRO 495)
  • UK price: from £2,359 (64 GB, 2 TB), as checked 17 October 2026. The 128 GB model is £3,519 and the 192 GB P495 is £6,679.
  • Where to buy: Minisforum UK. The 64 GB and 128 GB ship from a UK warehouse; the P495 ships from Hong Kong.
  • Best for: developers and home-lab users who want 128 GB or more of AI memory plus fast networking and room to add a card.
  • Our rating (desk review, not hands-on): ★★★½☆ 3.5/5

On this page

Which AI models can the Minisforum MS-S1 MAX run?

The 128 GB model holds models up to about 150 billion parameters at 4-bit, and the P495 about 260 billion, by our maths of about 0.6 GB per billion parameters.

Memory optionBiggest model at 4-bitExample modelsPublished speed
64 GB (about 48 GB for the GPU)About 70B, with short conversations onlyQwen3 32B (19.4 GB), Llama 3.3 70B (42.4 GB)No published figure for this model
128 GB (up to 96 GB for the GPU)About 150BLlama 3.3 70B at 8-bit (about 74 GB), gpt-oss-120b, Qwen3 Next 80BNo single-machine token figure published. Minisforum says two units together ran a 235B Q4 model at 10.87 tokens a second (maker claim)
192 GB P495 (up to 160 GB for the GPU)About 260BQwen3 235B-A22B (142.5 GB), DeepSeek V4.1 Flash Q2On the same chip and memory in Framework's tests: DeepSeek V4.1 Flash Q2, 14.0 tokens a second; MiMo V2.6 Flash, 18 tokens a second (Framework)

Allow 2-4 GB more for a long conversation. Mixture-of-experts models read only part of their weights for each word, so they run several times faster than dense models of the same size. That is what makes the P495 interesting: the big Qwen3 235B-A22B needs 142.5 GB at 4-bit, which fits in 160 GB but not in 96 GB.

How fast is it?

Reviewers have praised the MS-S1 MAX's speed but have not published its token rates. ServeTheHome found it the fastest of the four Ryzen AI Max+ 395 machines it tested, in both GPU work (MLPerf Client) and multi-core CPU work, where it led the Framework Desktop by 15%. It showed Qwen3 Next 80B running but gave no figure.

The chip is the same as in other Ryzen AI Max+ 395 machines, so their published figures are a fair guide:

  • Qwen3.8 27B, UD-Q4_K_XL, llama.cpp: 11.59 tokens a second (ROCm) and 12.09 (Vulkan), from Framework's August 2026 tests with a 2,048-token prompt.
  • Qwen3.6 35B-A3B (mixture of experts), UD-Q4_K_XL, llama.cpp ROCm: 52.13 tokens a second, from the same tests.
  • gpt-oss 120B in LM Studio: 31.41 tokens a second on the Beelink GTR9 Pro and about 45 on the Framework Desktop, both measured by ServeTheHome.

For the P495, Framework's tests of the same Ryzen AI Max+ PRO 495 with 192 GB give DeepSeek V4.1 Flash Q2 at 14.0 tokens a second in the DwarfStar engine (16.3 with DSpark speculative decoding) and GLM-5.3 Flash Q2 at 12. Minisforum says two P495 units clustered ran Qwen3.5 397B at 16 tokens a second. Framework expects only a minor speed gain from the PRO 495's 6.7% extra memory bandwidth; the real gain is room for bigger models.

Which version should you buy?

Buy the 128 GB model with a 2 TB drive at £3,519. It ships from the UK in days, gives the GPU 96 GB, and runs everything up to 70B dense and the fast 100B-plus mixture-of-experts models.

Skip the 64 GB model at £2,359 unless the dual 10 Gb and PCIe slot matter more to you than memory. The 64 GB Framework Desktop costs less. Buy the P495 only if you need models above about 150B in one box, such as Qwen3 235B-A22B at 4-bit. It costs almost double, ships from Hong Kong and may bring import VAT. GMKtec's EVO-X5 Pro has the same chip and 192 GB for £5,299 before VAT, and Framework's 192 GB DIY Desktop is £6,799 on pre-order with VAT included.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
MS-S1 MAX Local AI Pilot Edition64 GB2 TB£2,359Minisforum UK, UK warehouse
MS-S1 MAX128 GB2 TB£3,519Minisforum UK, UK warehouse
MS-S1 MAX128 GB4 TB£3,799Minisforum UK, UK warehouse
MS-S1 MAX-P495192 GB2 TB£6,679Minisforum UK, ships from Hong Kong

Minisforum shows these as sale prices against £2,949, £4,399, £4,749 and £8,349. It offers a 0% APR payment plan and takes bulk orders by email. Its refurbished store does not sell to the UK.

Hidden costs

  • Import charges on the P495: it ships from Hong Kong and Minisforum warns that customs charges may apply. At 20% import VAT that could be about £1,336, plus a courier fee. Minisforum sells optional tariff insurance that refunds the tax. Customs duty on computers is 0%.
  • Memory you cannot add later: all versions are soldered, so choose the size you will need.
  • Slower second drive slot: the second M.2 slot runs at PCIe 4.0 x1, so keep your models on the first drive.
  • Power: Notebookcheck measured up to 204 W at full load. ServeTheHome saw about 120 W running AI inference.
  • Screen and keyboard: not included.

Software: Ollama, LM Studio and the rest

The MS-S1 MAX ships with Windows 11 Pro. There is no NVIDIA CUDA and no Apple MLX; AI apps run through llama.cpp on AMD's ROCm or the Vulkan graphics interface.

  • llama.cpp: both ROCm and Vulkan backends work on this chip. Framework says the PRO 495 uses the same CPU and GPU designs, so the software carries over.
  • Ollama: on Linux it supports the Ryzen AI Max+ 395 through ROCm 7. On Windows it uses Vulkan, which is on by default. See our Ollama guide.
  • LM Studio: offers both a Vulkan and a ROCm llama.cpp runtime on Windows. See our LM Studio guide.
  • Setup step: set graphics memory in the UEFI or in AMD Software: Adrenalin Edition (Performance, then Tuning, then Variable Graphics Memory). On Linux, AMD recommends a small BIOS reservation and raising the shared limit with its amd-ttm tool, on kernel 6.18.4 or later, Fedora 43 or Ubuntu 26.04.
  • Network drivers: the Realtek 10 Gb chips are new, so a fresh Windows or Linux install may need Realtek's driver.

Noise, power and size

  • 222.1 x 206.3 x 77.1 mm and 2.8 kg, with a built-in 320 W power supply. It stands upright, lies flat or mounts in a rack.
  • Four modes: Performance (130 W sustained, 160 W peak on the 395), Balanced, Quiet and Rack. ServeTheHome measured 9-12 W at idle and up to 43 dBA at 1 m under load in Performance mode.
  • Notebookcheck found it too loud at idle at first (46 dB(A)); BIOS 1.03 cut that to 35.5 dB(A). Under load it measured 49.7-51.7 dB(A) at 15 cm. Idle power was 7.2-10.2 W.
  • For the P495, Minisforum quotes 32-37 dB at idle and 41.6-48 dB at full load at 50 cm, depending on mode.

Who makes the Minisforum MS-S1 MAX?

  • Minisforum was founded in 2018 by a group of computer engineers. Its Chinese name is Mingfan, and it also runs the Atomman and Kodlix brands.
  • The UK shop is run by Minisforum UK, with a UK warehouse for in-stock models and a returns address in Oldbury, West Midlands.
  • Pre-sales and after-sales help is by email to uk@minisforum.com; technical support is at support@minisforum.com.
  • Products bought from the UK shop have a two-year warranty.

What's good

  • UK warehouse stock for the 64 GB and 128 GB, with delivery in days.
  • The fastest Ryzen AI Max+ 395 system in ServeTheHome's tests.
  • Dual 10 Gb Ethernet, two 80 Gbps USB4 v2 ports and an internal PCIe x16 slot (x4 speed).
  • A 192 GB option with up to 160 GB for the GPU.
  • Built-in power supply and a slide-out chassis that opens with two screws.

What's not so good

  • Louder than the Framework Desktop and Beelink GTR9 Pro in ServeTheHome's tests, and Notebookcheck marked it down for noise.
  • The P495 ships from Hong Kong, with possible import VAT at the door.
  • Unclear returns: the shop advertises a 30-day money-back guarantee, but its written policy gives 7 days for unwanted items.
  • Dearer than the GMKtec EVO-X2 for the same chip and memory.
  • The second M.2 slot is only PCIe 4.0 x1.

Who should buy it?

Buy the MS-S1 MAX if you want a 128 GB AMD machine from UK stock with the best networking and expansion of the group, or one of the few 192 GB boxes you can order today. It suits developers who will add a network card or cluster several units. If you want quiet, the Framework Desktop is the better choice. If you want the lowest price, the GMKtec EVO-X2 is cheaper.

Buying in the UK

  • Where it comes from: the 64 GB and 128 GB models ship from Minisforum's UK warehouse, with delivery quoted at 3-4 days on the product page and 4-7 working days in the FAQ. The P495 ships from Hong Kong by FedEx or DHL in 7-10 working days.
  • VAT and duty: the UK shop does not say on the product page whether prices include VAT. GOV.UK says prices aimed at the public should include it, and UK warehouse stock has no import charges. For the Hong Kong P495, GOV.UK rules mean the courier may collect 20% VAT on goods over £135; duty on computers is 0%.
  • Warranty: two years from Minisforum, starting on the purchase date for stock items and the shipping date for pre-orders. Contact Minisforum before sending anything back.
  • Returns: the written policy allows refunds on unused, undamaged items within 7 days, with you paying return postage. Minisforum pays the postage for faults reported within 7 days.

Interesting facts about the Minisforum MS-S1 MAX

  • Minisforum says four MS-S1 MAX units clustered together ran DeepSeek-R1 671B at Q4_0, a 380 GB model.
  • The 80 Gbps USB4 v2 ports come from a separate controller chip, because the AMD processor does not support USB4 v2 itself, as ServeTheHome explains.
  • Its Rack mode is the only mode with no fan speed limit, for use in a server cabinet.

Alternatives to the Minisforum MS-S1 MAX

  • GMKtec EVO-X2 - pick it for the lowest price on 128 GB, if you accept import VAT and a one-year warranty.
  • Framework Desktop - pick it for quieter running, a UK company and 30-day returns, with a 192 GB option too.
  • NVIDIA DGX Spark - pick it if your tools need NVIDIA's CUDA software.

Our rating

AreaScore
Memory for AI5/5
Speed3/5
Value3/5
Software3/5
UK buying4/5
Overall★★★½☆ 3.5/5

The MS-S1 MAX is the best-equipped AMD AI mini PC you can buy from UK stock, and the P495 takes it to 192 GB. It costs more than rivals and runs louder.

Frequently asked questions

Can the Minisforum MS-S1 MAX run Llama 70B?

Yes. With 96 GB for the GPU, the 128 GB model holds Llama 3.3 70B at 4-bit (42.4 GB) or 8-bit (about 74 GB). Expect speeds like other Ryzen AI Max+ 395 machines, such as 3.7-3.8 tokens a second at Q6_K on the GMKtec EVO-X2.

What is the difference between the MS-S1 MAX and the MS-S1 MAX-P495?

The P495 uses the Ryzen AI Max+ PRO 495 with 192 GB of faster LPDDR5X-8533 and lets up to 160 GB go to the GPU, against 96 GB on the 128 GB model. It costs £6,679 against £3,519 and ships from Hong Kong rather than the UK.

Is the Minisforum MS-S1 MAX good for AI?

Yes, for running large models privately. Its memory lets it hold models no consumer graphics card can, though dense 70B models run at only a few tokens a second.

Sources

Scroll to Top