Can 4x RTX 3090 run DeepSeek V4.1 Flash?
Last updated:
Probably only partly: the weights are larger than its memory, so part of the model likely ran from system RAM. Measured 39.3 tok/s generation with part of the model offloaded to system RAM (llama.cpp PR #30817, GPU + CPU offload, UD-Q4_K_XL (GGUF); source).
Every measured result for DeepSeek V4.1 Flash on 4x RTX 3090
| Build | Quant / setup | Engine | tok/s | |
|---|---|---|---|---|
| 4x RTX 3090 + EPYC 9575F + 768 GB DDR5: DeepSeek V4.1 Flash UD-Q4_K_XL, 20.6 tok/s prose, 39.3 code | UD-Q4_K_XL (GGUF) | llama.cpp PR #30817, GPU + CPU offload | 39.3 CPU offload | source |
| 4x RTX 3090 + EPYC 9575F + 768 GB DDR5: DeepSeek V4.1 Flash UD-Q4_K_XL, 20.6 tok/s prose, 39.3 code | UD-Q4_K_XL (GGUF) | llama.cpp PR #30817, GPU + CPU offload | 20.6 CPU offload | source |
Memory: 4x RTX 3090 has 96 GB. DeepSeek V4.1 Flash weights are about 511 GB in its native format (vLLM recipe: 552B backbone + 196.6B Engram tables, 16B active per output token; official MXFP4/FP8 checkpoint ~511 GB). 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 V4.1 Flash: best home rig for DeepSeek V4.1 Flash.
DeepSeek V4.1 Flash on other hardware: 2x DGX Spark (GB10) · 2x Tesla V100 32GB · 4x DGX Spark (GB10) · Apple M3 Ultra · RTX 5090
Other models measured on 4x RTX 3090: Qwen3 235B-A22B · Qwen3.8 Flash-Next
Not enough VRAM? Rent before you buy. Run DeepSeek V4.1 Flash 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.