Local LLM Inference

Entry 12GB CUDA Build

Lowest-cost sensible CUDA entry for local AI. Good for 7B quantized chat, embeddings, and learning Ollama or llama.cpp; 13B models need tight quantization and shorter context.

€2,064

7 market-priced parts, 1 reference estimates

AMD 16GB ROCm Value Build

Value-focused AMD inference build with enough VRAM for useful 13B work and some 20B quantized experiments. Best when the target stack supports ROCm; choose NVIDIA if CUDA-only libraries are required.

€1,906

6 market-priced parts, 2 reference estimates

Estonian Value 16GB Build

Value build selected around parts that are easier to source from Estonian retailers. Good first serious local AI machine for 7B-13B models, RAG, and coding assistants without overbuying flagship hardware.

€2,099

4 market-priced parts, 4 reference estimates

Efficient 12GB CUDA Workstation

Efficient CUDA build for 7B-13B inference, coding assistants, and private chat without excessive heat or power draw. The 12GB VRAM is the limiter; 20B-class models require aggressive quantization and short context.

€2,337

6 market-priced parts, 2 reference estimates

Power-Efficient RTX 4000 Ada Build

Quiet, efficient always-on inference box with a 20GB professional NVIDIA GPU. Best for homelab serving, private assistants, and low-noise office use where power draw matters more than peak gaming performance; larger models still need conservative context settings.

€2,589

7 market-priced parts, 1 reference estimates

Blackwell 5070 Ti 16GB Build

Latest-generation 16GB NVIDIA option for buyers who want Blackwell features, GDDR7 bandwidth, and CUDA compatibility. Good for 13B-class inference; 30B-class models are experimental with tight context or offload and this is still not a replacement for 24GB+ VRAM builds.

€3,099

5 market-priced parts, 3 reference estimates

Radeon 24GB ROCm Compatibility Build

Good VRAM-per-euro only after the exact ROCm/PyTorch/Ollama stack is validated for the buyer's workload. CUDA-first tools, training recipes, and plugins should be assumed NVIDIA-first unless tested.

€3,422

6 market-priced parts, 2 reference estimates

Balanced NVIDIA 16GB

Balanced CUDA choice for local chat, coding assistants, embeddings, and 13B-class models. Some 30B-34B quantized workloads are experimental with tight context or CPU offload, but 16GB VRAM is the main limiter.

€3,647

6 market-priced parts, 2 reference estimates

24GB CUDA Inference Workstation

Strong consumer CUDA box for 13B-34B models, coding assistants, embeddings, and larger offload experiments. 70B-class use requires careful quantization, context settings, and realistic throughput expectations because the GPU has 24GB VRAM.

€6,203

5 market-priced parts, 3 reference estimates