Can RTX 3090 run DeepSeek R1 671B?
Probably only partly: the weights are larger than its memory, so part of the model likely ran from system RAM. Measured 8.97 tok/s generation with part of the model offloaded to system RAM (llama.cpp CUDA, Q2_K; source).
Every measured result for DeepSeek R1 671B on RTX 3090
| Build | Quant / setup | Engine | tok/s | |
|---|---|---|---|---|
| EPYC + single RTX 3090 (hybrid) | Q2_K | llama.cpp CUDA | 8.97 CPU offload | source ⚠ unchecked |
Memory: RTX 3090 has 24 GB. DeepSeek R1 671B weights are about 410.5 GB at Q4_K_M (671B params × 4.89 bits ÷ 8, before context/KV cache). Arithmetic, not a measurement; long contexts need more. Check other quants in the VRAM calculator.
Speeds are decode (generation) tokens/sec as reported by each source; quant, engine and context differ, so small gaps are noise.
All measured speeds for DeepSeek R1 671B: best home rig for DeepSeek R1 671B.
DeepSeek R1 671B on other hardware: Apple M3 Ultra
Other models measured on RTX 3090: Flint 40B-A2.5B · GLM 4.7 Flash · Qwen3.6 27B · Qwen3.6 35B-A3B · Qwen3.8 27B · gpt-oss 20B
Not enough VRAM? Rent before you buy. Run DeepSeek R1 671B on a bigger GPU on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.