Rigsmith

Build for… /

Build for AI

Running AI models like Llama, Qwen, Mistral or DeepSeek privately on your own PC with Ollama or LM Studio.

Budget

Goal: Models up to ~20B parameters

$1,368
  • Runs models up to about 22 billion parameters entirely on the graphics card (4-bit)
  • Replies about 1.6× as fast as a typical entry gaming PC (Ryzen 5 7600 + RTX 5060) for models that fit
  • 32 GB of memory: enough for most loading models
Processor
6C/12T
Graphics
16 GB card
Memory
32 GB DDR4
Storage
1 TB
Peak power
~308 W

Mid-range

Best value

Goal: Faster replies, ~20B models

$3,052
  • Runs models up to about 22 billion parameters entirely on the graphics card (4-bit)
  • Replies about 2.0× as fast as a typical entry gaming PC (Ryzen 5 7600 + RTX 5060) for models that fit
  • 64 GB of memory: plenty for loading models
Processor
14C/20T
Graphics
16 GB card
Memory
64 GB DDR5
Storage
2 TB
Peak power
~533 W

High-end

Goal: Models up to ~50B parameters

$5,992
  • Runs models up to about 50 billion parameters entirely on the graphics card (4-bit)
  • Replies about 3.8× as fast as a typical entry gaming PC (Ryzen 5 7600 + RTX 5060) for models that fit
  • 64 GB of memory: plenty for loading models
Processor
8C/16T
Graphics
32 GB card
Memory
64 GB DDR5
Storage
2 TB
Peak power
~715 W

Second-hand value pick

Used card

A second-hand RTX 3090 (24 GB) runs ~35B models, more than any new card under $2,000.

$3,154
  • Runs models up to about 35 billion parameters entirely on the graphics card (4-bit)
  • Replies about 2.2× as fast as a typical entry gaming PC (Ryzen 5 7600 + RTX 5060) for models that fit
  • 64 GB of memory: plenty for loading models
Processor
14C/20T
Graphics
24 GB card
Memory
64 GB DDR5
Storage
2 TB
Peak power
~583 W

Buying second-hand? Check the card works under load before paying, and use the used PC checker for whole PCs.

What matters

  • Video memory above all: it decides how big a model you can run at full speed.
  • An NVIDIA card: almost every AI tool supports NVIDIA (CUDA) first.
  • 32–64 GB of system memory for loading models and running bigger ones partly on the processor.

What to skip

  • A fast processor: it barely matters once the model is on the graphics card.
  • Cards with 8 GB of video memory: they limit you to small models.

Good to know

  • Models are big downloads (5–40 GB each), so a 2 TB drive fills up quickly.
  • AMD cards work for chatbots on Linux (ROCm) and in LM Studio, with less support elsewhere.
BUDGET_FINDERPro

The best build for your budget

For Local chatbots (LLMs): set exactly what you want to spend and we find the complete, compatible build that suits it best at that price.

Sign in and upgrade to use the budget finder

Each build is the cheapest complete, compatible PC from current parts that meets its goal, priced at the best Australian store prices as of 12/11/2026. Speed comparisons are Rigsmith estimates, not benchmarks. Open a build in the builder to swap parts or buy second-hand.