United States
PC builds for Qwen2.5 72B
A minimum build, the one we'd recommend, and a max-performance build — every part chosen for technical fit and checked for compatibility.
Hardware requirement
EstimatedRunning Qwen2.5 72B as Q4_K_M in Ollama at ~8,192 tokens of context.
How this is calculated
|
Model weights
derived
|
42.2 GB | 72.706 billion parameters × 0.58 bytes/param (Q4_K_M) ≈ 42.2 GB. |
|
KV cache (context memory)
derived
|
2.7 GB | 8,192 tokens × 80 layers × 2 × 8 KV heads × 128 head dim × 2 bytes ≈ 2.7 GB. Grows with context length. |
|
Runtime overhead
estimate
|
1 GB | Fixed VRAM overhead for Ollama (activations, CUDA context, fragmentation headroom). |
|
Total VRAM to run fully on GPU
estimate
|
45.9 GB | Weights + KV cache + overhead. This is what you need to keep the whole model on the graphics card for full speed. |
|
Recommended system RAM
estimate
|
64 GB | Headroom to load the model, run the OS, and hold any layers offloaded from the GPU. Rounded up to a standard kit size. |
|
Storage
estimate
|
2 TB NVMe recommended | Local model files are large (often tens of GB each). An NVMe SSD keeps load times reasonable. |
Memory figures are estimates derived from published model facts using a fixed method. They are not benchmarks. We do not publish tokens-per-second figures because we have no measured source for them.
Recommended builds
Minimum
"It runs."
Runs Qwen2.5 72B (Q4_K_M) in Ollama with GPU + system-RAM offload at ~8K context.
-
Graphics CardGeForce RTX 5090 Founders Edition$2,000
No single consumer GPU here holds this model fully (~46 GB needed). The GeForce RTX 5090 Founders Edition (32 GB) is the largest option; the rest runs from system RAM.
-
ProcessorAMD Ryzen 5 5600X$130
Local AI inference runs on the GPU, so a capable budget CPU like the AMD Ryzen 5 5600X keeps costs down without holding the model back.
-
MotherboardGigabyte B550M DS3H$100
Matches the CPU socket (AM4) and DDR4 memory; a solid value board with room to expand.
-
$87
This model wants about 64 GB of system RAM; the 32 GB DDR4 kit covers it with headroom for the OS and offload.
-
StorageCrucial P3 Plus$62
1 TB of NVMe — local model files are large (tens of GB each), and NVMe keeps load times short.
-
CPU CoolerCooler Master Hyper 212 Black Edition$30
Keeps the CPU cool and quiet under sustained load; compatible with the chosen socket.
-
Power SupplyCorsair HX1200 (2023)$250
Estimated system draw is about 753 W; the 1200 W unit leaves headroom for GPU power spikes and efficiency.
-
$40
Fits the motherboard form factor and has clearance for the graphics card (360 mm).
The model is larger than the 32 GB on this GPU, so some layers run from system RAM (slower but functional). A larger-VRAM GPU removes this.
Recommended
Your pick"The build we would buy."
Runs Qwen2.5 72B (Q4_K_M) in Ollama with GPU + system-RAM offload at ~8K context.
-
Graphics CardGeForce RTX 5090 Founders Edition$2,000
No single consumer GPU here holds this model fully (~46 GB needed). The GeForce RTX 5090 Founders Edition (32 GB) is the largest option; the rest runs from system RAM.
-
ProcessorAMD Ryzen 5 7600X$201
The AMD Ryzen 5 7600X is a modern DDR5 platform — plenty for feeding the GPU, prompt processing, and everyday work, with an upgrade path.
-
MotherboardGigabyte B650M AORUS ELITE AX$180
Matches the CPU socket (AM5) and DDR5 memory; a solid value board with room to expand.
-
MemoryG.Skill Trident Z5 RGB 64 GB$225
This model wants about 64 GB of system RAM; the 64 GB DDR5 kit covers it with headroom for the OS and offload.
-
StorageSamsung 990 Pro$170
2 TB of NVMe — local model files are large (tens of GB each), and NVMe keeps load times short.
-
CPU CoolerThermalright Peerless Assassin 120 SE$35
Keeps the CPU cool and quiet under sustained load; compatible with the chosen socket.
-
Power SupplyCorsair HX1200 (2023)$250
Estimated system draw is about 800 W; the 1200 W unit leaves headroom for GPU power spikes and efficiency.
-
$105
Fits the motherboard form factor and has clearance for the graphics card (360 mm).
The model is larger than the 32 GB on this GPU, so some layers run from system RAM (slower but functional). A larger-VRAM GPU removes this.
Max performance
"For users who value performance over budget."
Runs Qwen2.5 72B (Q4_K_M) in Ollama with GPU + system-RAM offload at ~8K context.
-
Graphics CardGeForce RTX 5090 Founders Edition$2,000
The most VRAM available here (32 GB) — maximum context and headroom, well above the ~46 GB this model needs.
-
ProcessorAMD Ryzen 7 9800X3D$472
The AMD Ryzen 7 9800X3D handles heavy multitasking, data prep, and CPU-side tokenization alongside a top GPU.
-
MotherboardGigabyte B650M AORUS ELITE AX$180
Matches the CPU socket (AM5) and DDR5 memory; a solid value board with room to expand.
-
MemoryG.Skill Trident Z5 RGB 64 GB$225
This model wants about 64 GB of system RAM; the 64 GB DDR5 kit covers it with headroom for the OS and offload.
-
StorageSamsung 990 Pro$170
2 TB of NVMe — local model files are large (tens of GB each), and NVMe keeps load times short.
-
CPU CoolerARCTIC Liquid Freezer III 360 A-RGB$150
Keeps the CPU cool and quiet under sustained load; compatible with the chosen socket.
-
Power SupplyCorsair HX1200 (2023)$250
Estimated system draw is about 825 W; the 1200 W unit leaves headroom for GPU power spikes and efficiency.
-
CaseNZXT H9 Flow (2023)$165
Fits the motherboard form factor and has clearance for the graphics card (435 mm).
The model is larger than the 32 GB on this GPU, so some layers run from system RAM (slower but functional). A larger-VRAM GPU removes this.
Totals are the sum of current reference prices for each part, not live retailer pricing. Component choices are made purely on technical fit. We may earn a commission when you buy through links on this site. It never changes the price you pay, and it never affects which hardware we recommend. Learn more.