Hardware for local models

Honest sizing: what runs where, without the benchmark theater.

Updated August 2026

The honest version

• Small models run acceptably on a modern laptop — fine for quick drafting and questions.
• Mid-size models want a dedicated GPU with generous VRAM; this is where most serious solo work lands.
• Large models need serious hardware — multi-GPU workstations or a server. This is the territory where a managed deployment starts making sense.

Rules of thumb

VRAM is usually the binding constraint: the model must fit in it (or be quantized down until it does) for GPU-speed responses. When a model does not fit, runtimes quietly fall back toward CPU and everything gets slow — if responses crawl, check that first.

When to stop buying hardware

If the hardware conversation is getting longer than the work conversation, ask about managed options — dedicated hardware Alluqi runs for you. See Managed deployment options.