At £1500 for chat, a real build is achievable: Qwen2.5-32B-Instruct Q4_K_M for around £403 — the best chat value at this budget.
Used office tower + Tesla P40 24GB
£340–£465
runs Qwen2.5-32B-Instruct Q4_K_M
no published benchmark yet
Qwen2.5-32B-Instruct Q4_K_M for around £403 — the best chat value at this budget
- This is a noisy, fiddly, enthusiast build — not for someone who wants it to 'just work'
- Corporate towers (Dell OptiPlex, HP EliteDesk/ProDesk) often ship with proprietary or low-wattage PSUs — check for a free PCIe x16 slot, physical case clearance, and budget for a replacement 750W PSU before buying a GPU for it
- Passive cooling only — no fan. You must fit a 3D-printed shroud and a high-static-pressure fan or it will thermal-throttle
- No video output at all — you need a second GPU or the host's onboard/iGPU video for a display
- Requires 'Above 4G Decoding' / Re-Size BAR enabled in the host BIOS to be recognised correctly
- Pascal architecture with no Tensor Cores — markedly slower than a same-VRAM RTX card, so pin your driver/CUDA versions and avoid bleeding-edge llama.cpp builds that assume newer GPUs
- Draws ~250W continuously under load — a real electricity cost if left on
- Buy new, not used, and confirm it has enough spare 8-pin PCIe power connectors for the GPU you're pairing it with — a failing or underpowered PSU can take a used GPU down with it
What's in it
Prices checked: 2026-08-26
How to put it together
- Enable 'Above 4G Decoding' / Re-Size BAR in the motherboard BIOS before installing the card
- 3D-print (or buy printed) a P40 fan shroud and fit a high-static-pressure 40-92mm fan to it
- Fit the 750W PSU and connect the two 8-pin EPS/PCIe power leads the P40 needs (check your specific card's power connector — some use CPU EPS connectors via an adapter)
- Seat the P40 in the primary PCIe x16 slot; keep the onboard/iGPU video output connected for the display
- Install the datacentre driver branch for Pascal and pin that driver version — avoid unattended driver upgrades
- Install llama.cpp with CUDA support and load a Q4_K_M model to confirm GPU offload and check temperatures under load
- Confirm the fan shroud fan is powered from a fan header or molex before first boot — running the card without airflow for even a few minutes under load risks thermal throttling or damage
Used office tower + RTX 3090
£950–£1120
runs Qwen2.5-32B-Instruct Q4_K_M
no published benchmark yet
Qwen2.5-32B-Instruct Q4_K_M for around £1035 — a strong chat option at this budget
- Noisy under load
- Corporate towers (Dell OptiPlex, HP EliteDesk/ProDesk) often ship with proprietary or low-wattage PSUs — check for a free PCIe x16 slot, physical case clearance, and budget for a replacement 750W PSU before buying a GPU for it
- Needs 2×8-pin PCIe power — check the host PSU before buying
- No warranty on used cards; test with a memory stress tool on arrival
- 2026 DRAM/GDDR shortage has pushed used prices to roughly double the 2024-25 norm, and a 3090 alone now costs more than a tower and PSU combined — shop around, don't assume older price guides still apply, and don't expect this to be a budget tier
- Buy new, not used, and confirm it has enough spare 8-pin PCIe power connectors for the GPU you're pairing it with — a failing or underpowered PSU can take a used GPU down with it
What's in it
Prices checked: 2026-08-26
How to put it together
- Open the tower side panel and remove the existing PSU
- Fit the 750W PSU and connect motherboard 24-pin and CPU 8-pin
- Seat the 3090 in the top PCIe x16 slot and connect both 8-pin PCIe cables
- Boot, install drivers, then run a llama.cpp model to confirm GPU offload
- Many office towers have short cases — check card length (~313mm) fits before buying
Mac mini M1 16GB
£280–£380
runs Llama 3.1 8B Instruct Q4_K_M
~12 tok/s (source)
Llama 3.1 8B Instruct Q4_K_M at ~12 tok/s for around £330 — a strong chat option at this budget
- Silent, low-power, and needs no assembly
- 10GB headroom only fits 7-8B-class coding models at roughly 12 tok/s on this class of hardware — too slow and too weak to be a pleasant coding assistant, so this build is not recommended for coding
- 16GB is unified memory shared between macOS and the model — expect roughly 10GB usable for a model before the system starts swapping
- No user-upgradeable RAM or storage — buy the 16GB configuration outright
What's in it
Prices checked: 2026-08-26
How to put it together
- Unbox the Mac mini and connect power, a display (HDMI or USB-C), keyboard and mouse
- Complete macOS setup and update to the latest supported macOS version
- Install Ollama for macOS (or build llama.cpp with Metal support) from the terminal
- Pull and run a Q4_K_M model sized to fit comfortably under ~10GB and confirm Metal (GPU) acceleration is active
- Close other memory-hungry apps before running a model — macOS will start swapping well before you hit the full 16GB
Mini PC — Ryzen 7 7840HS 32GB (CPU-only)
£320–£420
runs Llama 3.1 8B Instruct Q4_K_M
~12 tok/s (source)
Llama 3.1 8B Instruct Q4_K_M at ~12 tok/s for around £370 — a strong chat option at this budget
- CPU-only inference is fine for chat but too slow to be pleasant for coding assistance
- Roughly 24GB of the 32GB RAM is realistically usable for a model once the OS and other software are accounted for
- Silent and very low power (under 65W)
- No discrete GPU — inference runs on CPU (with limited iGPU/Vulkan offload); check the BIOS UMA/VRAM allocation setting if trying iGPU offload
- DDR5 dual-channel bandwidth is the real bottleneck for CPU inference speed, not the CPU core count
What's in it
Prices checked: 2026-08-26
How to put it together
- Unbox the mini PC, connect power and a display, and install/update the OS
- In the BIOS, check the UMA/iGPU memory allocation setting if you plan to try Vulkan iGPU offload
- Install llama.cpp (CPU build) or Ollama
- Pull a Q4_K_M model sized to fit within your usable RAM and confirm it runs, expecting single-digit-to-low-teens tokens/sec
- Don't buy this tier expecting coding-assistant speeds — it is a chat/summarisation box, not a coding rig