Local LLM workstations

Run Llama, Qwen, Gemma and gpt-oss on a machine you own. Your prompts and documents stay on it.

Runs Ollama · LM Studio · Open WebUI

  • From $3,999
  • GPU memory 16–96 GB
  • Ships in 7–10 business days
Novatech AI workstation in a black SilverStone SETA H2 case with a mesh front

Which models fit

Model 16 GBRTX 5070 Ti · RTX 5080From $3,999 24 GBRTX PRO 4000From $7,999 32 GBRTX 5090From $11,499 96 GBRTX PRO 6000From $28,799
Llama 3.1 8BMeta · 8B8-bitFullFullFull
Qwen3 14BAlibaba · 14B4-bit8-bitFullFull
gpt-oss-20bOpenAI · 21BFitsFitsFitsFits
Gemma 3 27BGoogle · 27BDoesn't fit in GPU memory4-bit8-bitFull
Qwen3 32BAlibaba · 32BDoesn't fit in GPU memory4-bit4-bitFull
Llama 3.3 70BMeta · 70BDoesn't fit in GPU memoryDoesn't fit in GPU memoryDoesn't fit in GPU memory8-bit
gpt-oss-120bOpenAI · 117BDoesn't fit in GPU memoryDoesn't fit in GPU memoryDoesn't fit in GPU memoryFits
  • Full16-bit, full quality
  • FitsRuns as released
  • 8-bitCompressed, close to full quality
  • 4-bitCompressed, small quality drop
  • Doesn't fit in GPU memory

Approximate GPU memory for the model plus a working context. Full means 16-bit. A model that doesn't fit can run partly on the CPU, much more slowly.

Check any model in the Build Finder

Every build comes with

  • Ready on first bootWindows 11 Pro or Ubuntu. Ollama, Open WebUI or ComfyUI set up on request.
  • 48-hour burn-in report
  • 3-year limited warranty
  • Lifetime technical support
  • Assembled in USA from new parts
  • Built in 7–10 business days, shipped free
How much GPU memory do I need to run an LLM locally?

About 6 GB for an 8B model at 4-bit, 22 GB for a 32B model and 44 GB for a 70B model, plus a few GB for long prompts. The Model Finder shows the build for any model.

Can it run a 70B model?

Yes. The RTX PRO 6000 build's 96 GB holds a 70B model at 8-bit. The 32 GB builds can run one partly from system memory, much more slowly.

Do my prompts or files leave the machine?

No. The model runs on the machine's own GPU, so prompts and documents aren't sent to an AI provider. Once a model is downloaded, it works offline.

Which software comes installed?

Windows 11 Pro, or Ubuntu if you ask. We can also set up Ollama and Open WebUI before it ships, so it's ready to use on first boot.

How fast will it run my model?

Speed depends on the model, its precision and how long your prompts are. Ask and we'll run your model on the build and send you the numbers before you order.

Your AI systems expert

  • POs and purchase cards welcome
  • Net terms
  • University & government procurement
Liquid-cooled processor and memory inside a Novatech AI server
What do you need?
What will you run?Pick all that apply.
How many people will use it?
How are you buying?Pick all that apply.
What's your budget?
Paste your specs
How many systems?
How do we reach you?
No obligation
A real engineer reviews it
Reply within 1 business day

WHAT HAPPENS NEXT

  1. An expert reads it.

    Your models, budget, and preferences, reviewed by the person who would build the system.

  2. You get a real spec.

    Configuration with pricing you can hand to procurement, in one business day.

  3. You decide.

    No obligation. If a smaller build serves your models better, that is what we recommend.