How Much VRAM to Run DeepSeek R1?
DeepSeek R1 is a 671B MoE model. At FP16, it requires ~670GB VRAM — a 8× H100 cluster. For most users, run the distilled variants: R1 Distill Qwen 32B (fits in 24GB at Q4) or R1 Distill Llama 70B (fits in 44GB at Q4). Full R1 via API is $0.55/M input — cheaper than self-hosting at most scales.
VRAM requirements by quantization level
| Quantization | VRAM needed | Quality | Fits on |
|---|---|---|---|
| FP16 (full) | 670GB | Full quality | 8× H100 80GB · 16× A100 80GB |
| Q4 (full) | 380GB | Good | 5× H100 80GB · 8× A100 80GB |
| R1 Distill Qwen 32B (Q4) | 22GB | Strong (distilled) | RTX 4090 · RTX 5090 · A100 40GB |
| R1 Distill Llama 70B (Q4) | 44GB | Excellent (distilled) | A100 80GB · H100 PCIe |
GPU recommendations & rental prices
Verified GPU rental prices across all major providers. Start in minutes, pay by the hour.
Frequently Asked Questions
Should I self-host DeepSeek R1 or use the API?
For full R1, use the API — self-hosting requires an H100 cluster ($26,000+/month on RunPod) which only makes sense at enormous scale. For distilled variants, self-hosting on a single RTX 4090 or A100 is practical and cost-effective at moderate inference volumes.
How does R1 Distill quality compare to full R1?
R1 Distill Llama 70B retains about 90% of full R1's reasoning capability on MATH-500 and AIME. R1 Distill Qwen 32B retains about 80%. For most practical reasoning tasks, the 70B distill is the right choice — it's open, self-hostable, and nearly as capable.
Data verified 2026-08-20. GPU rental prices change frequently — verify before purchasing. See methodology.