Models / DeepSeek V4.1 Flash / VRAM
How much VRAM do you need for DeepSeek V4.1 Flash?
Last updated:
Weight math from the model card, plus every measured RigBase result for DeepSeek V4.1 Flash. Leave extra room for context (KV cache).
Weight size by quant
| Quant | Bits/weight | Weights only |
|---|---|---|
| Native release format | – | ≈511 GB |
| Q8_0 | 8.50 | 795.5 GB |
| Q6_K | 6.56 | 614.2 GB |
| Q5_K_M | 5.70 | 533.7 GB |
| Q4_K_M | 4.89 | 458.0 GB |
| Q3_K_M | 4.00 | 373.9 GB |
| Q2_K | 3.16 | 295.6 GB |
Bits-per-weight for GGUF rows: llama.cpp quantize README. Parameter count: vLLM recipe: 552B backbone + 196.6B Engram tables, 16B active per output token; . We advise staying at Q4 or above for chat quality — see Is Q4 good enough?.
Measured speeds for DeepSeek V4.1 Flash
| Hardware | Quant | Engine | tok/s |
|---|---|---|---|
| 4x Nvidia GB10 (unified, DGX Spark) · 512 GB | EXL3 3.5 bpw + DSpark | TensorFold TP4, 4 nodes, 1 stream | 134.8 ⚡ |
| 4x Nvidia GB10 (unified, DGX Spark) · 512 GB | EXL3 3.5 bpw + DSpark | TensorFold TP4, 4 nodes, 1 stream | 78 ⚡ |
| 2x Nvidia GB10 (unified, DGX Spark) · 256 GB | 1017 checkpoint, EXL3 2.9-bit + DSpark | TensorFold Zig, 2 nodes, 1 stream | 74.9 ⚡ |
| 4x Nvidia GB10 (unified, DGX Spark) · 512 GB | 1017 checkpoint, FP8 + MXFP4 + DSpark | TensorFold Zig, 4 nodes, 1 stream | 65.8 ⚡ |
| 2x Nvidia GB10 (unified, DGX Spark) · 256 GB | EXL3 2.9-bit | vLLM, 2 nodes, 1 stream | 31.6 |
| Apple M3 Ultra (80-core GPU, unified) · 512 GB | not stated | oMLX v0.7.1.dev1, MTP enabled | 30.8 ⚡ |
| RTX 5090 · 32 GB | Maya-S (experts 2-3 bit), partly streamed from SSD | Project Maya 1.1.21 (maya-ds4) | 24.9 |
| 2x Tesla V100 32 GB (PG500-216, sm_70) · 64 GB | GSQ-RCO 2-bit | bnk v0.5.0, experts split VRAM/RAM (CPU | 13 |
| RTX 5090 32GB · 32 GB | MXFP4 experts + Q8_0/BF16 dense + fp8 engram table | llama.cpp fork dsv41-porte, expert strea | 8.9 |
| RTX 5090 32GB · 32 GB | MXFP4 experts (lossless repack) + Q8_0/BF16 dense | llama.cpp fork dsv41-porte, expert strea | 5.1 |
⚡ = speculative decoding (MTP, DFlash…). Full list: best home rig for DeepSeek V4.1 Flash.
Related
All DeepSeek V4.1 Flash builds VRAM + KV calculator By VRAM size GGUF vs AWQ vs FP8 vs NVFP4 How much system RAM?
FAQ
How much VRAM do you need for DeepSeek V4.1 Flash?
Weights alone are about 511 GB in the native release format. Add memory for the KV cache (context) and the runtime. The smallest-memory measured result in RigBase is 32 GB (RTX 5090) at 24.9 tok/s. Going below Q4 usually hurts quality for chat; see the Q4 guide.
What is the fastest measured speed for DeepSeek V4.1 Flash?
2x Nvidia GB10 (unified, DGX Spark): 31.6 tok/s plain decode (EXL3 2.9-bit, vLLM, 2 nodes, 1 stream).
Does system RAM replace VRAM?
Only for overflow. When the whole model fits in GPU or unified memory, system RAM barely matters for decode speed. See the system-RAM guide.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.