← AI Hashrate 中文

Popular GPU × Model Combinations

Ten-second gut check: pick a combo and see if it fits. Estimates use 8K context, Q4 quantization — switch context/quant on the fit page.

Entry

8.2B · 8GB VRAM

Entry GPU — can it run Qwen3?

8.2B · 8GB VRAM

New budget card vs Qwen3-8B

Mid-range

RTX 4070 Super 12GB

Qwen2.5-14B-Instruct

14.8B · 12GB VRAM

Mid-range workhorse — 14B Q4?

Qwen2.5-32B-Instruct

32.8B · 24GB VRAM

Used king — can 24GB fit 32B Q4?

RTX 3090 Ti 24GB

Qwen2.5-32B-Instruct

32.8B · 24GB VRAM

24GB ceiling — 32B fit test

High-end

8.2B · 24GB VRAM

Most popular GPU — Qwen3 FP16

27.0B · 24GB VRAM

Can 24GB run 27B Qwen3.5 Q4?

DeepSeek-R1-Distill-Qwen-32B

32.8B · 24GB VRAM

DeepSeek reasoning on a 4090?

27.0B · 16GB VRAM

New Blackwell — 27B fit?

27.8B · 32GB VRAM

32GB flagship — Qwen3.6 FP16?

Llama-3.3-70B-Instruct

70.6B · 32GB VRAM

70B Llama on 32GB? Q4 to the rescue

Apple Silicon

Mac mini M4 Pro 64GB

27.0B · 64GB VRAM

M4 Pro 64GB — Qwen3.5 FP16?

Mac mini M4 Pro 64GB

Llama-3.3-70B-Instruct

70.6B · 64GB VRAM

Apple unified memory — 70B Q4?

Mac Studio M3 Ultra 96GB

27.8B · 96GB VRAM

Mac Studio — max Apple LLM perf

Pro/Data Center

Qwen2.5-72B-Instruct

72.7B · 80GB VRAM

Data center — 72B FP16 fits

H200 141GB SXM5

27.8B · 141GB VRAM

141GB HBM — max context

RTX 6000 Ada 48GB

27.0B · 48GB VRAM

48GB pro card — 27B headroom