Framework Desktop for local AI: a quiet build-it-yourself AI desktop (2026)

By /
Framework Desktop for local AI: chip, memory and UK prices

The Framework Desktop is a 4.5-litre PC you finish yourself, built on AMD's Ryzen AI Max+ 395 with up to 128 GB of memory, and now also sold with 192 GB. Buy the 128 GB DIY Edition at £3,449, or about £3,720 with a drive, fan, cable and tiles. Up to 96 GB goes to the GPU, and Framework measured gpt-oss-120b at around 40-48 tokens a second. Prices include VAT. We rate it 4/5. See our guide to 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.

Framework Desktop at a glance for local AI: £1,959 for the 64 GB DIY Edition, up to 192 GB of memory, runs 70B models on 128 GB, rated 4 out of 5
Framework Desktop at a glance. Prices from UK stores, checked 17 October 2026.

Key facts

  • What it is: A small desktop PC from Framework, sold in the UK mainly as a DIY Edition that you finish yourself. It uses a standard Mini-ITX mainboard, a FlexATX power supply and a 120 mm fan.
  • Chip: AMD Ryzen AI Max 385 (32 GB model), Ryzen AI Max+ 395 with Radeon 8060S graphics and 40 compute units (64 GB and 128 GB), or Ryzen AI Max+ PRO 495 (192 GB).
  • Memory options: 32, 64 or 128 GB of LPDDR5X-8000, or 192 GB of LPDDR5X-8533. All are soldered, so you cannot add more later.
  • Usable for AI: on Windows, up to 96 GB of the 128 GB can be set as dedicated graphics memory in AMD's Adrenalin software, and the GPU can address up to 112 GB in total. On Linux, Framework recommends letting the GPU grow into unused memory instead.
  • Memory bandwidth: 256 GB/s (32, 64 and 128 GB); about 273 GB/s (192 GB)
  • UK price: from £1,959 (64 GB DIY Edition, no drive), as checked 17 October 2026. The 128 GB DIY Edition is £3,449. VAT included.
  • Where to buy: direct from Framework UK (Framework Computer UK Ltd).
  • Best for: people who want a quiet AI box from a company with a UK entity, VAT-inclusive prices, 30-day returns and a two-year warranty, and who do not mind fitting a fan and a drive.
  • Our rating (desk review, not hands-on): ★★★★☆ 4/5

On this page

Which AI models can the Framework Desktop run?

The 128 GB model holds models up to about 150 billion parameters at 4-bit, and the 192 GB model about 260 billion, by our maths of about 0.6 GB per billion parameters.

Memory optionBiggest model at 4-bitExample modelsPublished speed
32 GB, Max 385 (about 24 GB for the GPU)About 32BQwen3.8 27B (17.6 GB at UD-Q4_K_XL, Framework's pick), Qwen3 14B (8.5 GB)No published figure for this model
64 GB (about 48 GB for the GPU on Windows)About 70B, with short conversations onlyLlama 3.3 70B (42.4 GB); Framework fits Qwen3.5-122B-A10B at 3-bit (52.5 GB) on LinuxNo published figure for this model
128 GB (96 GB dedicated, up to 112 GB addressable)About 150Bgpt-oss-120b, Llama 3.3 70B at 8-bit (about 74 GB), DeepSeek V4 Flash 0731 at 2-bit (90.9 GB)gpt-oss-120b in LM Studio: about 40 tokens a second on Windows, 48 on Linux (Framework); DeepSeek V4 Flash 0731, llama.cpp ROCm: 14.90 (Framework)
192 GB, PRO 495 (up to 160 GB as dedicated graphics memory, more on Linux)About 260BQwen3 235B-A22B (142.5 GB), DeepSeek V4.1 Flash Q2 (about 163 GB loaded)DeepSeek V4.1 Flash Q2, DwarfStar: 14.0 tokens a second; MiMo V2.6 Flash, llama.cpp: 18 (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: in Framework's tests Qwen3.6 35B-A3B ran at 52.13 tokens a second against 11.59 for the dense Qwen3.8 27B. The big Qwen3 235B-A22B (142.5 GB at 4-bit) needs the 192 GB model.

How fast is it?

Framework publishes its own detailed, measured figures for this machine. Its August 2026 Linux tests used llama.cpp, a 2,048-token prompt and 128 tokens generated:

  • Qwen3.8 27B, UD-Q4_K_XL: 11.59 tokens a second with ROCm 7.14 and 12.09 with Vulkan. A Strix Halo-specific llama.cpp fork with its own 4-bit format reached 13.78.
  • Qwen3.6 35B-A3B, UD-Q4_K_XL, ROCm: 52.13 tokens a second, reading the prompt at 1,344.89 tokens a second.
  • DeepSeek V4 Flash 0731, UD-IQ2_XXS: 14.90 tokens a second (ROCm) and 14.70 (Vulkan).
  • With speculative decoding, Qwen3.8 27B rose to 27.15 tokens a second and Qwen3.6 35B-A3B to 92.30, both on Vulkan.

Framework's earlier LM Studio guide gave gpt-oss-120b at about 40 tokens a second on Windows and 48 on Linux, and Mistral Small 3.2 24B at Q6_K at about 10 and 12. ServeTheHome measured gpt-oss-120b (MXFP4) in LM Studio at about 45 tokens a second. That compares with 31.41 on the Beelink GTR9 Pro, also measured by ServeTheHome out of the box. The Minisforum MS-S1 MAX beat the Framework in ServeTheHome's CPU and Geekbench AI tests, at the cost of more noise.

Which version should you buy?

Buy the 128 GB DIY Edition at £3,449. It gives the GPU 96 GB, which holds 70B models at 8-bit and the fast 100B-plus mixture-of-experts models.

Framework's latest local AI guide argues the 64 GB model (£1,959) is now worth considering, because new 27B and 35B models run well on it. Choose it if you will mainly run models of that size. The 32 GB model is out of stock in the UK. The 192 GB PRO 495 at £6,799 is only worth it if you need models above about 150B in one box; it is on pre-order, with the second batch shipping in November.

UK prices and where to buy

VersionMemoryStorageUK priceWhere to buy
DIY Edition, Ryzen AI Max 38532 GBNone (add your own)£1,269, out of stockFramework UK
DIY Edition, Ryzen AI Max+ 39564 GBNone (add your own)£1,959Framework UK
DIY Edition, Ryzen AI Max+ 395128 GBNone (add your own)£3,449Framework UK
DIY Edition, Ryzen AI Max+ PRO 495192 GBNone (add your own)£6,799, pre-order, ships NovemberFramework UK
Pre-built, Ryzen AI Max+ PRO 495, Fedora 44192 GB2 TB£7,449, pre-order, ships NovemberFramework UK

All prices include VAT. The pre-built version is not offered with the Ryzen AI Max+ 395 in the UK. Pre-orders need only a deposit, which you can cancel for a full refund any time before shipment. No discounts were shown when I checked.

Hidden costs

  • Parts the DIY Edition leaves out: a 1 TB drive (£215 from Framework), a CPU fan (£19-£29), a UK power cable (£5) and 21 front tiles (£30 in black). That takes the 128 GB model to about £3,718.
  • Windows: £120 for Home or £220 for Pro if you want it; Linux is free.
  • Memory you cannot add later: all versions are soldered.
  • Power: ServeTheHome measured 140-170 W under load.
  • Screen and keyboard: not included. There are no import charges, as Framework's UK company sells with VAT included.

Software: Ollama, LM Studio and the rest

The DIY Edition comes with no operating system. Framework says Linux is the smoother choice for ROCm work, and in its December 2025 tests inference ran about 20% faster on Fedora 43 than on Windows 11. There is no NVIDIA CUDA and no Apple MLX; apps run through llama.cpp on AMD's ROCm or the Vulkan graphics interface.

  • llama.cpp: both ROCm and Vulkan backends work, and Framework's figures show neither is always faster. The ready-made Linux containers used in Framework's tests (AI Toolbox Cockpit) are available to download.
  • 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: Framework's own getting-started guide uses it on Windows. Install the Framework Desktop driver bundle, then set GPU offload to maximum. See our LM Studio guide.
  • Setup step: on Windows, set Variable Graphics Memory in AMD Software: Adrenalin Edition (Performance, then Tuning). 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.
  • The 192 GB model: Framework says the PRO 495 uses the same CPU and GPU designs, so the same software, including LM Studio and Ollama, carries over.

Noise, power and size

  • 96.8 x 205.5 x 226.1 mm (4.5 litres) and 3.1 kg, with a 400 W FlexATX power supply (80 Plus Gold at 110 V, Silver at 230 V).
  • The chip runs at 120 W sustained and 140 W boost. The Noctua fan option is rated at 28.8 dBA.
  • ServeTheHome measured 10-20 W at idle and 140-170 W under load. It ran the machine at 140 W for a day and a half without trouble.
  • With the Noctua fan it measured 36-37 dBA at idle and 39-41 dBA under load in a studio with a 34 dBA noise floor, quieter than the Minisforum MS-S1 MAX.

Who makes the Framework Desktop?

  • Framework Computer was founded in January 2020 by Nirav Patel and is based in San Francisco. It also makes the modular Framework Laptop 12, 13 and 16.
  • UK sales are made by Framework Computer UK Ltd, registered in England and Wales (company number 16391298), with a London address.
  • The Desktop was announced in February 2025, with pre-orders from 25 February and shipping from the third quarter of 2025.
  • Support is by email to support@frame.work, and Framework sells spare parts and publishes repair guides.

What's good

  • UK company, VAT-inclusive prices, a 30-day return window and a two-year warranty for UK buyers.
  • Very quiet: 39-41 dBA under load in ServeTheHome's test.
  • Framework publishes detailed, repeatable AI benchmarks and setup guides for this machine.
  • Standard Mini-ITX board, power supply and fan, so parts can be replaced or moved to another case.
  • 64 GB and 128 GB versions, plus a 192 GB option.

What's not so good

  • You must fit the fan, drive, side panel and tiles yourself. ServeTheHome's team fitted its fan the wrong way round at first.
  • Extra parts add about £270 to the headline price.
  • Only 5 Gb Ethernet and little room for expansion in the standard case.
  • No ready-built 128 GB version in the UK.
  • Dearer than the GMKtec EVO-X2 for the same chip and memory.

Who should buy it?

Buy the Framework Desktop if you want a quiet AI machine from a company with a UK entity, clear UK pricing and good published guidance, and you are happy to spend 10 minutes fitting parts. It suits Linux users most. If you want something ready to plug in, the Beelink GTR9 Pro comes built with a three-year warranty. If you want faster networking or a PCIe card, the Minisforum MS-S1 MAX is better equipped.

Buying in the UK

  • Who sells it: Framework Computer UK Ltd, a company registered in England and Wales, sells to UK addresses. Some islands and the Channel Islands may cost more or count as exports.
  • VAT and duty: Framework's UK terms say prices include VAT and the checkout total includes delivery and taxes. There is nothing more to pay on arrival.
  • Warranty: a two-year limited warranty from delivery for UK customers, on top of your rights under the Consumer Rights Act 2015. Framework may send replacement parts for you to fit.
  • Returns: you can return a product for any reason within 30 days of delivery. You pay the return postage unless it is faulty, and Framework refunds within 14 days of receiving it.

Interesting facts about the Framework Desktop

  • The front has 21 tile slots. You can buy colour and logo tiles or 3D print your own.
  • The PRO 495 version's PCIe x4 slot is open-ended, so it takes x8 and x16 cards, and Framework suggests 50 Gbps network cards for linking several machines.
  • Framework's local AI guide is written by Donato Capitella, who also works on DwarfStar, an inference engine built just for DeepSeek V4 Flash.

Alternatives to the Framework Desktop

  • Minisforum MS-S1 MAX - pick it for a factory-built box with dual 10 Gb Ethernet and a PCIe slot, if you accept more noise.
  • Beelink GTR9 Pro - pick it for a ready-built machine with VAT included and a three-year warranty.
  • GMKtec EVO-X2 - pick it for the lowest price on 128 GB, if you accept import VAT and a one-year warranty.

Our rating

AreaScore
Memory for AI5/5
Speed3/5
Value3/5
Software3/5
UK buying5/5
Overall★★★★☆ 4/5

The Framework Desktop is the easiest of the AMD AI boxes to buy in the UK and one of the quietest to own. You pay a little more and do a little assembly.

Frequently asked questions

Can the Framework Desktop run Llama 70B?

Yes. The 128 GB model holds Llama 3.3 70B at 4-bit (42.4 GB) or 8-bit (about 74 GB) in its 96 GB of graphics memory. The 64 GB model can just hold a 4-bit 70B, with little room for a long chat.

Should I buy the 64 GB or 128 GB Framework Desktop?

Buy 128 GB if you want 70B models or the 100B-plus mixture-of-experts models. The 64 GB model, £1,490 cheaper, is enough for the newer 27B and 35B models that Framework now recommends.

Is the Framework Desktop available in the UK?

Yes. Framework sells it direct through its UK company, with VAT included. Framework lists the 64 GB and 128 GB DIY Editions, and the 192 GB version is on pre-order for November.

Sources

More guides

Scroll to Top