Hardware Tiers

Three levels of local AI. Pick the power you need.

Choose Your Power

Every LlamaBox runs the same Rocky Linux operating system, the same OpenAI-compatible API, the same LAN-only security model. The only difference is GPU power — and the models you can run.

Starter

$10,000

Perfect for small teams and solo professionals.

  • RTX 4090 (24GB VRAM)
  • AMD Ryzen 9 9950X
  • 64GB DDR5 RAM
  • 2TB NVMe SSD
  • Qwen3.6-35B default model
  • Up to 3 concurrent users
  • llama.cpp inference engine
  • LAN-only endpoint
View Full Pricing

Pro

$20,000

Our recommended tier for serious businesses.

  • RTX 6000 Ada (48GB ECC)
  • AMD Ryzen 9 9950X
  • 128GB DDR5 RAM
  • 2TB NVMe SSD
  • Qwen3.6-70B default model
  • Up to 10 concurrent users
  • vLLM inference engine
  • LAN-only endpoint
View Full Pricing

Ultra

$40,000

Enterprise-grade. Maximum inference power.

  • RTX 6000 Blackwell (96GB GDDR7)
  • AMD Ryzen 9 9950X
  • 256GB DDR5 RAM
  • 2TB NVMe SSD
  • Qwen3.6-110B default model
  • Up to 20 concurrent users
  • vLLM inference engine
  • LAN-only endpoint
View Full Pricing

The Build

Premium components. Professional design. Built to last.

Component Specification
Case Lian Li O11D XL White — dual glass panels, showcases GPU
Motherboard ASUS ProArt X870E-CREATOR WIFI — 10Gb Ethernet, USB4, 4 M.2 slots
CPU AMD Ryzen 9 9950X — 16 cores / 32 threads
RAM G.Skill Flare X5 Neo DDR5-6000 CL30 (64GB/128GB/256GB by tier)
SSD Samsung 990 Pro 2TB NVMe — 1,200 TBW endurance
GPU RTX 4090 (24GB) / RTX 6000 Ada (48GB ECC) / RTX 6000 Blackwell (96GB)
PSU Seasonic PRIME TX-1000 Titanium — 1000W
Cooling Arctic Liquid Freezer III 360 — handles 230W CPU at ~$110
Fans 2x Lian Li Uni Fan SL140 White
OS Rocky Linux 9 — enterprise-grade, pre-configured, "set it and forget it"
Design Glass case design — GPU visible through tempered glass, "the engine of privacy"

Model Capabilities by Tier

Every tier runs Qwen3.6 models at optimal quantization for the GPU.

Tier VRAM Default Model Max Model Concurrent Users
Starter 24GB Qwen3.6-35B Q4-KS 70B Q2 Up to 3
Pro ⭐ 48GB ECC Qwen3.6-70B Q4-KS 110B Q2 Up to 10
Ultra 96GB GDDR7 Qwen3.6-70B Q8 110B Q4-KS Up to 20

* Q4-KS = 4-bit quantization, high quality. Q2 = 2-bit quantization, reduced quality. Q8 = 8-bit quantization, maximum quality.

Inference Engines

Different tiers use different engines, optimized for their hardware.

llama.cpp (Starter Tier)

Single binary, zero dependencies. Built-in chat web UI on same port. OpenAI-compatible API via --api flag. Perfect for 1-2 concurrent users. Simple, reliable, fast to deploy.

vLLM (Pro & Ultra Tiers)

PagedAttention + continuous batching for multi-user throughput. OpenAI-compatible API (native). Supports safetensors (AWQ-INT4, FP8) model formats. Handles 10-20 concurrent users effortlessly.

Ready to Order?

All tiers are fully customizable. Need more RAM? More storage? Different GPU? We'll build it.

View Pricing & Support