DeepSeek on NVIDIA Blackwell

Can I run DeepSeek-R1 671B-A37B on an NVIDIA DGX Spark (GB10)?

Not at Q4_K_M — it needs 380.1 GB against 96.0 GB available. You would need 4 of these cards.

Won't fit

380.1 GB of 96.0 GB · 396%
0395 GB
Weights 379.0 GB
KV cache 0.5 GB
Runtime overhead 0.6 GB
Over the limit 284.1 GB

Short by 284.1 GB. You can run it with 15 of 61 layers on the NVIDIA DGX Spark (GB10) and the rest in system RAM, at roughly 2.40 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation2.40tok/s
Prompt processing709tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of DeepSeek-R1 671B-A37B on a NVIDIA DGX Spark (GB10)

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 1249.8 GB 1250.9 GB Won't fit 0.64
INT8 / W8A8 8.50 663.6 GB 664.7 GB Won't fit 1.26
Q8_0 (GGUF) 8.50 663.6 GB 664.7 GB Won't fit 1.26
FP8 (E4M3) 8.00 624.9 GB 626.0 GB Won't fit 1.35
Q6_K 6.56 512.4 GB 513.5 GB Won't fit 1.69
Q5_K_M 5.67 442.9 GB 444.0 GB Won't fit 2.00
Q5_K_S 5.52 431.2 GB 432.3 GB Won't fit 2.05
Q4_K_M 4.85 379.0 GB 380.1 GB Won't fit 2.40
Q4_K_S 4.58 358.0 GB 359.1 GB Won't fit 2.57
Q4_0 4.55 355.6 GB 356.7 GB Won't fit 2.59
AWQ 4-bit 4.25 334.5 GB 335.6 GB Won't fit 2.79
GPTQ 4-bit 4.25 334.5 GB 335.6 GB Won't fit 2.79
MXFP4 4.25 334.5 GB 335.6 GB Won't fit 2.79
IQ4_XS 4.25 332.3 GB 333.4 GB Won't fit 2.81
Q3_K_M 3.91 305.8 GB 306.9 GB Won't fit 3.09
IQ3_M 3.70 289.4 GB 290.5 GB Won't fit 3.36
IQ3_XXS 3.06 239.6 GB 240.7 GB Won't fit 4.29
Q2_K 2.63 206.1 GB 207.2 GB Won't fit 5.30
IQ2_XXS 2.06 161.7 GB 162.8 GB Won't fit 7.58
IQ1_M 1.75 137.5 GB 138.6 GB Won't fit 10.2

Also worth checking