Lagoon

Questions, answered straight

How much memory do I need to run AI locally?

Memory is the deciding spec. 16GB runs 8–14B models, 24–48GB runs 20–32B models, and 64GB or more runs 70B models. Capacity decides which models fit at all; memory bandwidth (GB/s) decides how fast they answer.

What is the best budget computer for local AI?

The Mac mini M6 with 16GB, at about $899. It is silent, sips power, and runs Lagoon Micro's models — Qwen3-14B, Phi-4, and Granite 4.1 8B — all day. On the Windows side, the RTX 5060 Ti 16GB is the budget graphics card for a custom build.

Can a laptop run a 70B model?

Only a unified-memory laptop. No Windows laptop GPU holds more than 24GB of VRAM, which caps Windows laptops at 32B-class models. A MacBook Pro with M5 Max at 64GB, or the ASUS ROG Flow Z13 at 128GB, can run 70B models.

Is a Mac or a Windows PC better for local AI?

Macs win on capacity per dollar: Apple Silicon treats all of its memory as one pool, so a 128GB Mac Studio holds models no single graphics card can. Windows PCs with NVIDIA cards win on raw speed and CUDA software support at a given memory size, but consumer VRAM tops out at 32GB.

What is the best open-weight model to run locally right now?

For most workstations, Qwen3.8-27B: it needs about 18GB and sits on the Pareto frontier in our chart. On 16GB machines the strongest small models are Qwen3-14B, Phi-4, and Gemma 4 12B. For 70B-class intelligence, gpt-oss-120B and Qwen3-235B-A22B lead, if you have the memory.