Skip to content
All hardware
FrameworkAMD Strix HaloDesk tier · £2k – £7k

Framework Desktop

The community favourite — the best-documented local-LLM box on the platform, from the repairability company.

Price

~$2,851 (128 GB + 1 TB)

US verified Mar 2026 · UK GBP via configurator (~£2.2–2.5k expected) · rose ~$460 in the RAM crisis

verified 2026-07 · supply & lease options in every proposal

Framework Desktop — small-form-factor PC with a black grid front panel, three-quarter viewFramework
128 GB
Unified memory

LPDDR5X-8000, soldered

256 GB/s
Memory bandwidth

~212 GB/s measured

4.5 L
Mini-ITX chassis

front 120 mm fan

~£2.4k
Typical build

128 GB + 1 TB

The machine

Framework's 4.5-litre desktop is the Strix Halo machine the local-AI community standardised on: the benchmark grids, the Linux guides and the llama.cpp tuning threads mostly run on this box. Dual M.2, 5 GbE, Wi-Fi 7 and Framework's modular ethos (though the LPDDR5X is necessarily soldered).

Pricing has tracked the RAM crisis: the 128 GB board launched at $1,999 and rose ~$460 in the squeeze — ~$2,851 with 1 TB fitted at last verified snapshot. Framework sells UK-direct in GBP (configurator pricing; expect ~£2.2–2.5k) and has publicly committed to cutting prices when memory normalises.

vs its siblings: The de facto community reference machine for the whole platform.

Memory, to scale

128 GB model-visible · bandwidth is the speed limit

Unified LPDDR5X

128 GB · 256 GB/s

LPDDR5x

For scale

DGX Spark — 128 GB @ 273 GB/s

RTX PRO 6000 — 96 GB @ 1.79 TB/s

Mac Studio M3 Ultra — 512 GB @ 819 GB/s

DGX Station GB300 — 748 GB coherent

Capability

What it actually runs

Declared from research and benchmarks, not computed marketing — tokens-per-second figures are cited where a real measurement exists.

  • gpt-oss-120bMXFP4with headroom30–55 tok/s — the platform's best-documented result
  • Qwen3.5 122BQ4fits~65 GB — Linux VRAM window required
  • Qwen3.6 35BQ4with headroomfast triage-class serving
  • gpt-oss-20bMXFP4with headroomflies — high-volume pipeline work
Specification

The full sheet

Compute

APU
AMD Ryzen AI Max+ 395 — 16× Zen 5, Radeon 8060S (40 CU), XDNA 2 NPU

Memory

Unified memory
128 GB LPDDR5X-8000 (32/64 GB boards available)
VRAM allocation
~96 GB Windows · ~110 GB+ Linux

Storage & I/O

Storage
2× M.2 NVMe
Networking
5 GbE + Wi-Fi 7 · 2× USB4
PSU
500 W internal

Where it shines

  • The best community documentation of any box in this catalogue
  • Linux-first: the ~110 GB VRAM window is real here
  • Repairable, modular, honest company pricing communication
  • Cheapest credible 128 GB unified path after GMKtec

The trade-offs

  • 5 GbE, not 10 — weaker clustering story than Beelink's dual-10 GbE
  • DIY assembly (mainboard + case) — not for procurement teams
  • Same 256 GB/s physics as every Strix box
  • UK GBP price not independently verified at research time

Buy this box for

Engineers who want the reference community boxLinux-based local inference with maximum VRAM allocationBudget-conscious pilots that still need 120B-class models
The platform

Understanding AMD Strix Halo

Ryzen AI Max+ 395

AMD's Ryzen AI Max+ 395 ('Strix Halo') puts 16 Zen 5 cores, a 40-CU Radeon 8060S GPU and up to 128 GB of unified LPDDR5X-8000 on one package. Framework, HP, GMKtec, Beelink, Corsair and AMD itself all ship it as a compact desktop, from roughly £1,700 to £2,700 in the UK.

It is a mixture-of-experts machine: gpt-oss-120b runs at 30–55 tok/s and even Qwen3-235B squeezes in at 3-bit (~11 tok/s), but dense 70B models crawl at ~5 tok/s because 256 GB/s of bandwidth is the ceiling. The software stack (Vulkan / ROCm / llama.cpp / LM Studio) is genuinely usable in 2026 but remains a step behind CUDA for production serving and fine-tuning.

It is the right first box for proving local AI on real workloads before committing to bigger silicon — and the refresh is close: AMD's Ryzen AI Max 400 series ('Gorgon Halo', announced May 2026) lifts the ceiling to 192 GB with OEM systems from Q3 2026.

Sources & verification

Specifications and prices verified 2026-07 against the sources below. The memory shortage is repricing this market monthly — we re-verify at quote.

Compare against the rest of the catalogue, or have us spec it against your workloads.

Back to catalogue
Make it real

Spec the Framework Desktop for your workload

We'll size it against your real token volumes, quote supply and lease options, and stand up the models with evals that prove they do the job — or tell you honestly that a different box (or the API) wins.