Enthusiast local AI build
This tier in our build guides is for people who use local models every day and are tired of waiting. It is still a single-GPU machine, but built around the fastest consumer memory available rather than the cheapest path to a given capacity. You get large-context language model sessions, video generation that finishes in a reasonable time, and fine-tuning runs that complete overnight instead of over a weekend. For the broader question of what any given machine can handle, see the hardware pages.
Target
Roughly 24 to 32 GB of VRAM, or an Apple Silicon machine with 64 to 128 GB of unified memory.
Approximately, this class covers:
- Mid-sized to large language models at 4-bit quantization, or mid-sized models at higher precision when output quality matters more than parameter count.
- Long context windows, because the extra capacity goes to the KV cache rather than the weights, and cache is what long documents and long agent runs consume.
- Video diffusion work at usable resolutions and clip lengths, which is the first tier where video stops being an experiment.
- Full LoRA training on mid-sized models and quantized fine-tuning on larger ones.
Approximate, as always. A model that fits at short context may not fit at long context, and multimodal models carry extra memory costs beyond their weight files.
GPU or unified memory
Two cards define this tier:
- RTX 5090, 32 GB of GDDR7. The capacity step and the memory bandwidth step arrive together, and bandwidth is what sets generation speed once a model fits.
- RTX 4090, 24 GB of GDDR6X. Same capacity as the mid-range cards, noticeably more throughput, which is the difference between a working setup and a fast one.
An Apple Silicon machine makes sense at this tier only if capacity beats speed for you. A high-memory M3 Max configuration addresses far more model memory than any consumer card, so it loads models a 5090 cannot, and runs them slower than a 5090 would. If your work is loading big models occasionally, that trade is good. If it is running a model hard all day, it is not.
The rest of the box
- CPU. A high-core-count desktop CPU with enough PCIe lanes to feed the card at full width. Tokenisation, data preparation and multi-process serving all benefit; inference itself mostly does not.
- System RAM. 64 GB is the sensible floor and 128 GB is not excessive. Weights stage through system memory on load, fine-tuning datasets live there, and CPU offload lets you run something slightly oversized at reduced speed rather than not at all.
- Storage. Several terabytes of NVMe. At this tier you keep multiple quantizations of the same model, full-precision weights for fine-tuning, video frame output and checkpoints, and all of it wants to be on fast storage.
- Power and cooling. Generous supply headroom and a case designed for sustained load, not peak load. Plan the airflow around a card that will sit near its limit for hours. Undervolting is worth the afternoon it takes.
Upgrade path
The workstation tier is where you stop adding speed and start adding cards, which buys capacity for model classes no single consumer GPU holds. The mid-range tier gets you most of the same model classes for less machine, just slower.