Build guides
This index sits under /hardware. Each guide below describes a complete machine for running models locally: the memory class it targets, the kind of work it suits, and the parts of the system that matter beyond the GPU.
Individual cards get their own pages under GPUs for local AI. These guides are about the whole box, including system RAM, storage, power and cooling.
The four tiers
- Budget build: the cheapest honest entry point, what a small single card actually runs, and which corners are safe to cut.
- Mid range build: a single mainstream card in the common consumer memory class, suitable for everyday chat, coding help and image work.
- Enthusiast build: a top consumer card or a dual card setup, aimed at larger models, longer context and faster diffusion work.
- Workstation build: many cards or large unified memory, for fine-tuning, serving several users, or video generation.
Tiers are memory classes, not price brackets
These guides are deliberately not organised by spend. Component prices move constantly, and a number written down here would be wrong soon after and useless to anyone reading later. What does not move is the relationship between memory and capability.
A tier here means two things:
- A memory class. How much VRAM, or unified memory, the machine has in total. That figure sets which model sizes fit after quantization, how much room is left for the KV cache, and therefore how long a context window you can hold open.
- A use case. What you intend to do. Interactive chat, batch document processing, image generation, video generation and training all stress a machine differently, and a build that is generous for one is thin for another.
Read the tier that matches your workload, then use the current price of the cards in that memory class to decide whether to buy now. The guide tells you what to look for. Only you can see what the parts cost today.
Two further notes that apply across every tier. System RAM matters more than people expect, because models are loaded from disk through it and because inference can spill into it when a model does not quite fit. And fast storage matters when you keep many model files around, since a large weights file has to be read every time you switch.