Model finder
Which workstation runs your model?
Pick a model and a precision. You'll see how much GPU memory it needs and the most affordable Novatech build that runs it fully on the GPU.
- 17 models
- 4 GPU memory sizes
- From $3,999
Our builds by GPU memory
Pick a model to see the build that runs it.
At a glance
Every model, every build
Each model at its most common precision. A check means it runs fully on the GPU.
| Model | 16 GBRTX 5070 Tifrom $3,999 | 32 GBRTX 5090from $11,499 | 96 GBRTX PRO 6000from $28,799 | 192 GB2× RTX PRO 6000from $92,998 |
|---|---|---|---|---|
| Llama 3.1 8B8-bit · ~10 GB | ||||
| Qwen3 8B8-bit · ~10 GB | ||||
| Gemma 3 12B4-bit · ~11 GB | ||||
| Qwen3 14B4-bit · ~11 GB | ||||
| gpt-oss-20bAs released · ~14 GB | ||||
| Mistral Small 3.24-bit · ~17 GB | – | |||
| Gemma 3 27B4-bit · ~19 GB | – | |||
| Qwen3 Coder 30B-A3B4-bit · ~20 GB | – | |||
| Qwen3 32B4-bit · ~22 GB | – | |||
| DeepSeek-R1 Distill Llama 70B4-bit · ~44 GB | – | – | ||
| Llama 3.3 70B4-bit · ~44 GB | – | – | ||
| gpt-oss-120bAs released · ~66 GB | – | – | ||
| Qwen3 235B-A22B4-bit · ~130 GB | – | – | – | |
| Llama 3.1 405B4-bit · ~230 GB | Bigger than one build. Ask about a custom multi-GPU build | |||
| Stable Diffusion XLFull precision · ~12 GB | ||||
| FLUX.1 [dev]Full precision · ~28 GB | – | |||
| Whisper large-v3Full precision · ~5 GB | ||||
Approximate GPU memory for the model plus a working context. Longer context and more users need more. Models that don't fit can run partly on the CPU and system memory, much more slowly.
Sizing questions
What do 4-bit and 8-bit mean?
They're compressed versions of a model. 4-bit needs about a quarter of the memory of full precision with a small drop in quality, and 8-bit is close to full quality at about half the memory. Most people run large language models at 4-bit or 8-bit.
What happens if a model doesn't fit?
Tools like Ollama and llama.cpp can run part of the model on the CPU and system memory, but it runs much more slowly. For everyday use, pick a build where the model fits fully in GPU memory.
Why leave room to spare?
Long prompts, big documents and several people at once all need GPU memory on top of the model itself. A few GB of headroom keeps responses fast.
Can two GPUs share one model?
Yes. On a multi-GPU server, inference software like vLLM splits one model across the cards, so it can use their combined memory.
Not sure?
Tell us your model. We'll size the build.
Send us the model, the context length and how many people will use it. Our team will recommend a build, including multi-GPU options.