Best rigs by budget
These are shopping brackets, not a price feed. The only euro figure on this page is one builder's report that RigBase cites. Product links go through /go so a real affiliate URL can replace the manufacturer link later. Nothing here has a tracking ID.
Entry
8–12 GB cards people already own or buy used. Small models at Q8, and low-teens models at Q4 when the estimate fits.
- RTX 2060 Super8 GB vram
8 GB card from the 20-series. Fine for small models at Q4.
- RTX 3060 12GB12 GB vram
12 GB GDDR6. The 12 GB model is the one matched from the benchmark set.
- RTX 3080 Ti12 GB vram
12 GB GDDR6X. Same memory class as the 12 GB 3060, with more bandwidth.
Midrange
16 GB cards. A practical home for 14B-class models at Q4 and 8B models at Q8.
- RTX 4060 Ti 16GB16 GB vram
16 GB version of the 4060 Ti. Rows for the 8 GB card are not mixed in.
- RTX 4070 Ti12 GB vram
12 GB GDDR6X. The 4070 Ti Super (16 GB) is a different card and is not folded in.
- RTX 5060 Ti 16GB16 GB vram
16 GB Blackwell card. Mixed-GPU rows that pair it with another card are excluded.
- RTX 5070 Ti16 GB vram
16 GB card in NVIDIA's 50-series. Single-card rows only.
Flagship
24–32 GB single cards. Dense models around 27–32B fit at Q4. A 70B Q4 does not.
- RTX 309024 GB vram
24 GB GDDR6X. The usual used flagship in local-model builds.
- RTX 409024 GB vram
24 GB GDDR6X. Single-card rows only; multi-GPU rigs are separate entries.
- RTX 509032 GB vram
32 GB Blackwell flagship, from rows that report 32 GB on a single card.
- TITAN RTX24 GB vram
24 GB Turing card. Included because the benchmark set has single-card runs on it.
- Radeon RX 7900 XTX24 GB vram
24 GB RDNA 3 card. Runs in the dataset are mostly llama.cpp on ROCm or Vulkan.
- Radeon AI Pro R970032 GB vram
32 GB workstation Radeon. Only rows for a single card are matched.
- Intel Arc Pro B7032 GB vram
32 GB Arc Pro card. Dual-B70 rows are not matched to this single-card entry.
Multi-GPU
Several 24 GB cards in one box. This is the usual way a 70B Q4 gets a measured run.
- 2× RTX 309048 GB vram
Two 24 GB cards, 48 GB combined. Tensor-parallel runs in the dataset are kept on this entry.
RigBase's row for this class cites a builder who described two used 3090s as about 2000 EUR all-in. That is one report, not a current street price. Original post
- 4× RTX 309096 GB vram
Four 24 GB cards, 96 GB combined, from rows that say 4× RTX 3090 and 96 GB.
Unified memory
Apple silicon and Strix Halo. The operating system shares the pool, so the fit rule uses 70% of the stated memory.
- Ryzen AI Max+ 395 64GB64 GB unified
Strix Halo with 64 GB unified memory. The iGPU shares that pool with the OS.
- Ryzen AI Max+ 395 128GB128 GB unified
Strix Halo with 128 GB unified memory, from rows that report 128 GB on the 8060S iGPU.
- M1 Ultra 128GB128 GB unified
M1 Ultra with 128 GB unified memory. Mac Studio class in the benchmark notes.
- M2 Ultra 192GB192 GB unified
M2 Ultra, 76-core GPU, 192 GB unified memory.
- M3 Pro 18GB18 GB unified
M3 Pro at 18 GB unified memory, the figure the matched rows report.
- M3 Ultra 256GB256 GB unified
M3 Ultra rows that report 256 GB unified memory.
- M3 Ultra 512GB512 GB unified
M3 Ultra rows that report 512 GB unified memory.
- M4 16GB16 GB unified
Base M4 at 16 GB unified memory. The form factor is not stated in every source row.
- M4 32GB32 GB unified
Base M4 at 32 GB unified memory.
- M4 Pro 24GB24 GB unified
M4 Pro at 24 GB. Some source rows call the machine a MacBook Pro.
- M4 Pro 48GB48 GB unified
M4 Pro, 20-core GPU, 48 GB unified memory.
- M4 Max 36GB36 GB unified
M4 Max at 36 GB unified memory.
- M5 Max 36GB36 GB unified
M5 Max rows that report 36 GB unified memory.
- M5 Max 128GB128 GB unified
M5 Max rows that report 128 GB unified memory.
- M5 Ultra 96GB96 GB unified
M5 Ultra, 64-core GPU, 96 GB, described as a Mac Studio in the source rows.
- M5 Ultra 256GB256 GB unified
M5 Ultra, 80-core GPU, 256 GB, described as a Mac Studio in the source rows.
- M6 32GB32 GB unified
M6 with a 12-core GPU and 32 GB unified memory, as the matched rows describe it.
Workstation
96 GB workstation GPUs and GB10-class machines for higher precision and the larger MoE checkpoints.
- RTX Pro 6000 Blackwell96 GB vram
96 GB workstation card. Rows that also include a second GPU are not matched here.
- NVIDIA GB10128 GB unified
Grace Blackwell Superchip as reported with 128 GB unified memory. Multi-node rows are separate.
- 2× NVIDIA GB10256 GB unified
Two GB10 machines, 256 GB combined, from rows that say 2× Nvidia GB10.