Hardware for local AI, by tier
For local AI, memory matters more than raw speed: VRAM decides which models fit, and memory bandwidth decides how fast they talk.
8GB GPUs
The entry tier. Runs 4B-9B models fully on the GPU at Q4 quantisation, which covers everyday chat, summarising and light coding.
Runs: gemma4:e4b, qwen3.5:9b, ministral-3:8b…
12GB GPUs
The sweet spot for 8B-14B models with room for longer context. 12B models like Gemma 4 12B fit fully on the GPU.
Runs: gemma4:12b, qwen3.5:9b, ministral-3:14b…
16GB GPUs
Unlocks 20-24B models such as gpt-oss 20B, Mistral Small 3.2 and Devstral Small 2, plus 14B models with long context.
Runs: gpt-oss:20b, mistral-small3.2, devstral-small-2…
24GB+ GPUs
Runs the best local models: Qwen3.6 27B, Gemma 4 31B and the 30-35B mixture-of-experts models, fully on the GPU.
Runs: qwen3.6:27b, qwen3.6:27b-coding, gemma4:31b…
Mini PCs & unified memory
Machines where the GPU shares system memory. Slower per token than a big GPU, but they can hold far larger models in a small, quiet box.
Runs: gpt-oss:120b, qwen3.6:35b, gemma4:26b…
Disclosure: some product links on these pages may become affiliate links, marked rel="sponsored". We pick hardware on what it can run, not on commission.