All hardware
Every build in the database (73 total), grouped by category. Each links to a page showing which open-weights LLMs run on it and how fast — same picks you'd see in the live picker.
Sizing a build around one model? See Qwen 3.6 27B VRAM & GPU requirements — per-quant VRAM, which GPUs can run it, and measured tokens per second on the RTX 5090, RTX 3090, AMD R9700, Intel Arc Pro B70 and used MI50, then the cheapest build that fits. Cross-shopping unified-memory boxes? RTX Spark vs DGX Spark vs RTX 5090 covers capacity vs speed. Picking a llama.cpp backend for one of these cards? CUDA vs Vulkan for llama.cpp breaks it down by vendor. On Apple Silicon? Apple M5 Max for local LLMs covers what the 128GB MacBook Pro actually runs and whether to wait for the M5 Ultra Mac Studio.
Apple Silicon (15)
- Mac Mini M4 (16 GB)
$799 — the cheapest serious local LLM box.
- MacBook Air M4 (16 GB)
- Mac Mini M4 (24 GB)
Cheapest serious local LLM box on the planet.
- MacBook Pro M4 Pro 48 GB
- MacBook Pro M4 Max 64 GB
- MacBook Air M5 (16 GB)
- MacBook Pro M5 Pro 48 GB
- MacBook Pro M5 Max 64 GB
- MacBook Pro M5 Max 128 GB
- Mac Studio M4 Max 128 GB
The 70 B sweet spot at low watts.
- Mac Studio M4 Max 36 GB
Cheapest path into M4 Max silicon — fine for 30 B-class agents, no headroom above.
- Mac Studio M3 Ultra 96 GB
The accessible M3 Ultra — 70 B + 120 B-MoE at low watts.
- Mac Studio M3 Ultra 256 GB
- Mac Studio M3 Ultra 512 GB
- 2× Mac Studio M3 Ultra 512 GB cluster (TB5 / MLX)
1 TB unified Apple Silicon — used market only.
NVIDIA — consumer GPUs (11)
- RTX 3060 12 GB build
Entry-level CUDA on a budget.
- Single RTX 3090 (used) build
- Dual RTX 3090 (used) build
Best $/tps for 70 B locally.
- Quad RTX 3090 (used) build
- Single RTX 4090 build
- Single RTX 5090 build
Fastest single GPU under $5 K.
- Dual RTX 5090 build
- NVIDIA RTX Spark (128 GB)
Coming soon — DGX Spark silicon in Windows laptops and desktops, announced at Computex 2026.
- Tinybox Green (6× RTX 4090, 144 GB)
- Tinybox Green v2 (4× RTX 5090, 128 GB)
- Tinybox Pro (8× RTX 4090, 192 GB)
NVIDIA — pro workstation (20)
- Single Tesla P40 24 GB (used) build
Cheapest 24 GB CUDA card on the planet.
- Quad Tesla P40 (96 GB) homelab build
- Single Tesla P100 16 GB (used) build
- Tesla V100 32 GB SXM2 mod build
$200 datacenter card + a mod = 32 GB CUDA HBM2.
- NVIDIA DGX Spark (128 GB)
- 2× DGX Spark cluster (256 GB unified, CUDA)
Cheapest officially-supported CUDA cluster.
- 4× DGX Spark cluster (512 GB unified, CUDA)
- 8× DGX Spark cluster (1024 GB unified, CUDA)
Eight Sparks, one terabyte, no DGX-Cloud bill.
- MSI EdgeXpert MS-C931 (128 GB)
GB10 with 4 TB Gen5 storage at Amazon-Prime convenience.
- ASUS Ascent GX10 (128 GB)
- Dell Pro Max with GB10 (128 GB)
- Lenovo ThinkStation PGX (128 GB)
- Single RTX A6000 48 GB (Ampere) build
- Single RTX 6000 Ada 48 GB build
- Single RTX Pro 6000 Blackwell 96 GB build
The 70 B + 120 B sweet spot on one card.
- Single RTX Pro 5000 Blackwell 48 GB build
48 GB CUDA workstation card at half a 5090's wattage.
- Dual RTX Pro 6000 Blackwell build
- Quad RTX Pro 6000 Blackwell build (384 GB)
Workstation-form datacenter, four cards deep.
- 8× RTX Pro 6000 Blackwell server (768 GB)
Quarter-DGX-H200 price, two-thirds the memory.
- 12× RTX Pro 6000 Blackwell rack (1152 GB)
Datacenter-class memory pool on PCIe parts.
Datacenter (H100/H200/B200) (9)
- Single H100 80 GB workstation
- 8× H100 80 GB server
- Single H200 141 GB workstation
H100 silicon with 1.76× the memory.
- DGX H200 — 8× H200 server (1.13 TB HBM3e)
The 2025-26 reference frontier-LLM inference box.
- Single B200 180 GB workstation
Frontier silicon, single-card form.
- DGX B200 — 8× B200 server (1.44 TB HBM3e)
2026's frontier-LLM reference rig.
- Single AMD Instinct MI300X 192 GB workstation
- Single AMD Instinct MI325X 256 GB workstation
MI300X with a third more memory.
- Single AMD Instinct MI355X 288 GB workstation
AMD's first FP4 card — matches B200 on memory.
AMD Instinct + Radeon (11)
- AMD Ryzen AI Max+ 395 (128 GB)
Best $/GB unified memory.
- 2× Strix Halo cluster (256 GB unified)
- Single AMD Radeon Pro W7900 48 GB build
- Dual AMD Radeon Pro W7900 build
- Single AMD Radeon Pro W7800 32 GB build
- Single AMD Radeon AI Pro R9700 32 GB build
Cheapest new 32 GB workstation card.
- Dual AMD Radeon AI Pro R9700 build (64 GB)
- Single AMD Radeon RX 9070 XT 16 GB build
$599 RDNA4 — the consumer R9700.
- Single AMD Instinct MI50 32 GB (used) build
Best $/(GB·bandwidth) on Earth right now.
- Quad AMD MI50 32 GB (128 GB) homelab build
128 GB HBM2 for under $2 K.
- Tinybox Red (6× 7900 XTX, 144 GB)
Intel Arc Pro (5)
- Single Intel Arc Pro B70 build
Cheapest 32 GB discrete card on the market.
- Dual Intel Arc Pro B70 build
- Quad Intel Arc Pro B70 build
- Octuple Intel Arc Pro B70 cluster
- Single Intel Arc B580 12 GB build
Strix Halo / DGX Spark (2)
- 4× Strix Halo cluster (512 GB unified)
Cheapest 512 GB unified-memory rig on Earth.
- 8× Strix Halo cluster (1024 GB unified)
1 TB unified RAM for half the price of two Mac Studios.
Last updated 2026-07-11