Witbank running local LLMs build under R20,000

Quick Answer

A Witbank local-LLM build under R20,000 hinges on VRAM. An RTX 4060 Ti 16GB (around R10,500) with a Ryzen 5 7600 and 32GB DDR5 runs 7B and 13B models smoothly, and the full build stays under R20,000.

VRAM Determines Model Size

Running language models locally needs the model to fit in GPU memory. A 7B model in 4-bit quantisation needs about 6GB; a 13B model wants 10-12GB. The RTX 4060 Ti 16GB handles both comfortably with headroom. Under R20,000 it is the best VRAM-per-rand choice for local LLMs, beating faster but smaller-memory cards for this use.

The Rest Of The Build

Pair the GPU with a Ryzen 5 7600 (around R4,500), 32GB DDR5 (around R3,000) and a 1TB NVMe SSD (around R1,100) to store model files of 4-15GB each. A 650W 80+ Gold PSU covers it. Ollama and LM Studio download and run quantised 7B-13B models with a few clicks, making this an approachable home AI workstation.

Witbank Notes

Evetech ships assembled LLM rigs to Witbank and across Mpumalanga by courier. Inference runs offline, so after the initial model download you chat with the model without a connection - handy where bandwidth is metered. The warm inland climate suits a mesh case with two intake fans to keep the GPU under 75C during long generation runs, which protects token-generation speed over sustained use.

FAQ

What model sizes can I run under R20,000?

7B and 13B models run smoothly on the RTX 4060 Ti 16GB. Larger 30B+ models need 20GB+ of VRAM and a bigger budget.

Which LLM software is easiest?

Ollama and LM Studio are the simplest - they handle downloading and running quantised models with minimal setup. Both run well on this build.

Do I need 32GB or 64GB of system RAM?

32GB is enough for GPU-resident 7B-13B models. 64GB only helps if you offload larger models partly to the CPU, which runs slower.