Best DIY Hybrid MoE Overall
Custom EPYC 7443 Milan + RTX 3090 24GB + 128GB 8-Ch DDR4-3200
£2,834
Combines 24GB GDDR6X (936 GB/s) for attention/KV cache & fast prompt prefill with 128GB 8-channel DDR4-3200 ECC RDIMM (205 GB/s) for sparse MoE routed experts at 1/4th the cost of consumer DDR5.
Specs, Pros & Caveats (6)
Hybrid (New + Used) · 520W · 34 dBA (Noctua Tower)
- 128GB 8-channel DDR4-3200 ECC RDIMM costs £480 vs £2,199 for 2-channel DDR5
- RTX 3090 24GB delivers 936 GB/s VRAM bandwidth & fast CUDA FlashAttention prefill
- Runs DeepSeek V3/R1, Qwen3.5-122B-A10B & gpt-oss-120B via KTransformers / ik_llama
- Full PCIe 4.0 x16 lanes across 7 slots on Gigabyte MZ32-AR0 for future 2nd/3rd 3090
- Verify EPYC 7443 CPU is unlocked (not Dell/Lenovo PSB fuse-locked)
- Dense 70B models (>24GB) that spill to CPU RAM decode at ~11–13 tok/s (MoE models hit 26–42 tok/s)
800 GB/s Dense 70B King
Apple Mac Studio M1 Ultra (20C CPU / 48–64C GPU / 128GB Unified)
£2,450
Unmatched 800 GB/s unified memory bandwidth under £3,000. Beats Strix Halo by 2.8x on 70B dense token generation while drawing under 90W at whisper-quiet acoustics.
Specs, Pros & Caveats (6)
Refurbished / Used · 85W · 22 dBA (Near-silent)
- 800 GB/s memory bandwidth — fastest 70B dense decode (~19 tok/s) under £3,000
- 114GB allocatable to Metal/MLX via sysctl iogpu.wired_limit_mb
- CeX UK includes mandatory 5-Year Warranty (£2,450–£2,925 Grade B/C)
- Draws <90W under full LLM load; tiny footprint
- Slow prompt prefill (150–300 tok/s) on long >16k contexts due to M1 GPU compute & no SME
- No hardware bfloat16 (use Q4/Q5/Q8 MLX or FP16 weights to avoid bf16 emulation slowdown)
Best New Turnkey 128GB (UK Stock)
Bosgame M5 / GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB LPDDR5X-8000, 2TB)
£2,914
Brand-new x86/Linux Mini PC with 16 Zen 5 cores, 40-CU Radeon 8060S GPU, and 128GB LPDDR5X-8000. Unlocks up to 120GB usable VRAM on Linux 6.16+ via TTM page limits.
Specs, Pros & Caveats (6)
New · 120W · 38 dBA
- Up to 120GB usable GPU memory on Linux (options ttm pages_limit=31457280)
- 81–85 tok/s decode & 750 tok/s prefill on Qwen 3 30B-A3B MoE (Vulkan RADV / ROCm)
- Runs 120B MoE (gpt-oss-120b / DeepSeek V4 Flash) entirely in unified memory at 24–27 tok/s
- Standard x86_64 Fedora/Ubuntu compatibility + 120W wall power
- 256 GB/s bandwidth limits 70B dense models to ~14–16 tok/s (vs 19 tok/s on M1 Ultra)
- Price doubled in 2026 from £1,499 launch due to LPDDR5X shortage
Best Value New 128GB (£1,050 Under Cap)
GMKtec EVO-X2 / Bosgame M5 128GB (Direct EU/CN Hub Import)
£1,950
Identical Ryzen AI Max+ 395 128GB hardware sourced directly via vendor EU/CN logistics hubs, bypassing the ~£1,000 UK retail distributor markup.
Specs, Pros & Caveats (5)
New · 120W · 38 dBA
- Saves ~£960–£1,150 compared to Amazon UK / UK retail listings
- Leaves £1,050 budget headroom for 4TB NVMe upgrades or USB4/Oculink networking
- Full 120GB usable VRAM under Linux 6.16+ with Mesa RADV Vulkan
- Return-to-base vendor warranty rather than Amazon UK / CeX local returns
- Same 256 GB/s memory bandwidth ceiling on dense models
Best Sub-£2,300 Capacity (216GB Total)
Dell Precision 7920 Tower + RTX 3090 24GB + 192GB 12-Ch DDR4 ECC
£2,280
Refurbished dual-socket Dell enterprise tower with 1400W Platinum PSU, 192GB 12-channel DDR4 ECC RDIMM, and a used RTX 3090 24GB — £720 below the £3,000 ceiling.
Specs, Pros & Caveats (5)
Refurbished / Used · 560W · 36 dBA (Baffled Chassis)
- 192GB 12-channel DDR4 ECC + 24GB VRAM = 216GB total memory for £2,280
- Includes 1400W 80+ Platinum hot-swap PSU and acoustic-baffled workstation chassis
- Room and PSU headroom to drop in a second RTX 3090 later for 48GB VRAM
- Dual-socket NUMA requires pinning threads or running --numa distribute in llama.cpp
- Requires genuine Dell 8-pin GPU power harness (part D92C9 / GYF0C, ~£19)
Best All-New GPU Build (32GB GDDR7)
Dual RTX 5060 Ti 16GB (32GB GDDR7) + Ryzen 9 5900X + 128GB DDR4
£2,583
Pairs two Blackwell RTX 5060 Ti 16GB cards (32GB GDDR7 total, native FP4/FP8 Tensor Cores) with 128GB DDR4 (£814 vs £2,199 for DDR5) on a proven AM4 platform.
Specs, Pros & Caveats (5)
New · 430W · 33 dBA
- 100% brand-new retail components with full UK manufacturer warranties
- 32GB fast GDDR7 VRAM holds 30B–35B dense/MoE models & 36k+ KV cache entirely in VRAM
- 128GB DDR4 saves £1,385 vs DDR5, funding both Blackwell GPUs
- Models exceeding 32GB VRAM spill into 2-channel DDR4 (51 GB/s), slowing >70B decode
- Best for models <=35B in pure VRAM or sparse MoE with small active expert footprints
Best High-Clock 8-Channel Workstation
Lenovo ThinkStation P620 (Threadripper PRO 3955WX) + RTX 3090 + 128GB 8-Ch
£2,840
Single-socket 8-channel DDR4-3200 workstation with 4.3 GHz Zen 2 Threadripper PRO clocks and 128 PCIe 4.0 lanes, paired with a 24GB RTX 3090.
Specs, Pros & Caveats (4)
Refurbished / Used · 500W · 35 dBA
- Single NUMA domain with 8-channel DDR4-3200 (~205 GB/s) and high 4.3 GHz boost clocks
- 1000W standard ATX-pinout Platinum PSU included in Lenovo P620 chassis
- Zero motherboard assembly required — slot in 8 RDIMMs and the RTX 3090
- CPU is Lenovo PSB-locked (stays paired with the P620 motherboard)
Best In-Stock Apple Silicon (<£3k)
Apple MacBook Pro 16" M3 Max (16C CPU / 40C GPU / 96GB Unified)
£2,840
While 96GB (32GB shy of 128GB), it is the only high-memory Apple Silicon unit physically in stock under £3,000 at CeX UK today, adding native hardware bfloat16 and 2x faster prefill than M1 Ultra.
Specs, Pros & Caveats (5)
Refurbished / Used · 65W · 25 dBA
- Physically in stock at CeX UK for £2,840 with 5-Year UK Warranty
- Native hardware bfloat16 + dynamic caching + 2x faster prompt prefill than M1 Ultra
- Includes 16" Liquid Retina XDR display & battery UPS built in
- 96GB total unified memory (~86GB usable VRAM) rather than 128GB
- 400 GB/s bandwidth vs 800 GB/s on M1/M2 Ultra
Contiguous 32GB Single-GPU Build
Gigabyte Radeon AI PRO R9700 32GB + Ryzen 7 5700X + 128GB DDR4
£2,965
Uses AMD's 32GB RDNA 4 workstation blower card with 128GB DDR4 system RAM to fit 32GB models on a single card without multi-GPU tensor split overhead.
Specs, Pros & Caveats (4)
New · 390W · 37 dBA (Blower)
- Single contiguous 32GB VRAM pool (640 GB/s) — no PCIe peer-to-peer split needed
- Blower cooler exhausts heat directly out of the case
- ROCm software ecosystem requires more tuning than CUDA on new architectures
- 2x RTX 5060 Ti 16GB (£900) delivers the same 32GB VRAM for £820 less
128GB Pure HBM2 VRAM (Enthusiast)
4x NVIDIA Tesla V100 32GB (128GB HBM2) + EPYC 7402P + 128GB DDR4
£2,950
Four enterprise 32GB HBM2 GPUs on a 7-slot EPYC SP3 board give 128GB of 900 GB/s HBM2 VRAM plus 128GB 8-channel DDR4 RAM under £3,000.
Specs, Pros & Caveats (4)
Refurbished / Used · 1150W · 52 dBA (High Airflow Required)
- 128GB of pure 900 GB/s HBM2 GPU VRAM + 128GB 8-channel system RAM (256GB total)
- Runs 70B Q8 and 120B Q4 models 100% inside HBM2 VRAM at 28–38 tok/s
- Volta architecture (SM70) lacks native FlashAttention-2, bfloat16, and INT4/FP8 Tensor Cores
- 1,150W peak power draw and requires 3D-printed fan shrouds for passive server cards
Avoid — The 2026 DDR5 Trap
Consumer AM5 Ryzen 5 7600 + 128GB DDR5-5600 + Intel Arc B580 12GB
£2,953
Illustrates why standard consumer DDR5 DIY fails in Oct 2026: 128GB of 2-channel DDR5 costs £2,199 alone, leaving only £345 for a 12GB GPU and bottlenecking large models to 75 GB/s.
Specs, Pros & Caveats (3)
New · 280W · 31 dBA
- £2,199 spent on 128GB DDR5 that only delivers ~75 GB/s dual-channel bandwidth
- Only 12GB GPU VRAM; 70B models crawl at ~4.5 tok/s over 2-channel DDR5
Over £3k Cap (Board Alone £3,149)
Framework Desktop DIY (Ryzen AI Max+ 395, 128GB LPDDR5X-8000)
£3,538
Open-source modular 4.5L Mini-ITX Strix Halo workstation. After Framework's May 2026 +11% UK DRAM price hike, even the bare mainboard (£3,149) exceeds the £3,000 cap.
Specs, Pros & Caveats (3)
New · 120W · 32 dBA (120mm Noctua/CoolerMaster)
- Standard Mini-ITX mainboard + 120mm fan + PCIe x4 slot for 50GbE RoCE clustering
- Official Fedora support and reproducible Framework LLM benchmark suite
- Exceeds £3,000 budget (£3,149 mainboard only; £3,538 with case/PSU/SSD)
£350 Over Cap at CeX (£3,350)
Apple Mac Studio M2 Ultra (24C CPU / 60C GPU / 128GB Unified)
£3,350
Adds 20–30% faster prompt prefill and native bfloat16 over M1 Ultra while keeping 800 GB/s bandwidth, but starts at £3,350 used in the UK.
Specs, Pros & Caveats (2)
Refurbished / Used · 95W · 22 dBA
- 800 GB/s unified memory bandwidth + native hardware bfloat16
- Starts at £3,350 at CeX UK (£350 above the £3,000 ceiling)
£1,780 Over Cap After 2026 Price Hike
NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, 128GB Unified)
£4,780
1 PFLOP FP4 Blackwell tensor compute with 128GB unified memory and ConnectX networking, but 2026 DRAM price hikes pushed UK pricing to £4,780.
Specs, Pros & Caveats (4)
New · 170W · 36 dBA
- Full NVIDIA CUDA stack + native NVFP4 Tensor Cores (102 tok/s on Qwen3.5-35B-A3B)
- Fast prompt prefill (1,800+ pp/s) despite 273 GB/s memory bus
- £4,780 UK retail price is £1,780 over the £3,000 budget
- 273 GB/s bus still bottlenecks dense 70B Q4 decode to ~15–17 tok/s
£2,399 Over Cap (Apple RAM Tax)
Apple Mac Studio M5 Max (18C CPU / 40C GPU / 128GB Unified, 1TB)
£5,399
Apple's late-2026 M5 Max Mac Studio starts at £2,499 (36GB), but upgrading to 40-core GPU + 128GB unified memory adds £2,900, bringing the total to £5,399.
Specs, Pros & Caveats (3)
New · 105W · 22 dBA
- Massive prompt prefill speedup (>2,200 pp/s) via M5 Neural Accelerators & SME
- 546 GB/s unified memory in a silent 2.7kg desktop
- £5,399 UK price is nearly double the £3,000 budget ceiling