NVIDIA · unified x86

NVIDIA DGX Spark (GB10)

NVIDIA DGX Spark (GB10) has 128 GB of unified memory at 273 GB/s — about 119.04 GiB usable after driver and compositor overhead. 2091 of 2118 indexed models fit at 64K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
128 GB
LPDDR5X-8533
Bandwidth
273 GB/s
256-bit bus
Tensor FP16
125 TF
dense
TDP
140 W
$3999 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1796vision language 191audio tts 21image 2audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Llama-3_3-Nemotron-Super-49B-v1_5Q5_K_M49.9B32.96 GiB85.00 GiB118.90 GiB0.14 GiB2±12.9%
Valkyrie-49B-v2.1I1-Q5_K_M49.9B32.96 GiB85.00 GiB118.90 GiB0.14 GiB2±12.9%
Llama-3_3-Nemotron-Super-49B-v1Q5_K_M49.9B32.96 GiB85.00 GiB118.90 GiB0.14 GiB2±12.9%
command-a-plus-05-2026-bf16MoEIQ4_NL219B116.79 GiB1.29 GiB118.88 GiB0.16 GiB8±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB6.38 GiB118.83 GiB0.21 GiB7±37%
grok-2MoEUD-IQ3_XXS270B109.27 GiB8.50 GiB118.71 GiB0.33 GiB3±37%
Qwen3.5-397B-A17BMoEIQ2_XS403B116.79 GiB1.00 GiB118.64 GiB0.40 GiB11±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB1.00 GiB118.26 GiB0.78 GiB10±37%
step-3.5-flashIQ4_NL199B103.97 GiB13.30 GiB118.10 GiB0.94 GiB2±12.9%
dots.llm1.instMoEQ4_1143B84.24 GiB32.94 GiB118.00 GiB1.04 GiB3±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB3.98 GiB117.97 GiB1.07 GiB8±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB1.21 GiB117.76 GiB1.28 GiB10±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB16.73 GiB117.72 GiB1.32 GiB2±12.9%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB16.73 GiB117.72 GiB1.32 GiB2±12.9%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB1.00 GiB117.68 GiB1.36 GiB11±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ3_K_L236B113.97 GiB2.24 GiB117.05 GiB1.99 GiB9±37%
DeepSeek-V2.5MoEQ3_K_L236B113.97 GiB2.24 GiB117.05 GiB1.99 GiB9±37%
DeepSeek-Coder-V2-InstructMoEQ3_K_L236B113.97 GiB2.24 GiB117.05 GiB1.99 GiB9±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB8.23 GiB116.92 GiB2.12 GiB5±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB6.11 GiB116.33 GiB2.71 GiB6±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB6.11 GiB116.33 GiB2.71 GiB6±37%
Llama-3_1-Nemotron-51B-InstructQ4_151.5B30.18 GiB85.00 GiB116.12 GiB2.92 GiB2±12.9%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB7.44 GiB115.89 GiB3.15 GiB3±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ3_XXS269B102.71 GiB12.22 GiB115.77 GiB3.27 GiB5±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB8.23 GiB115.42 GiB3.62 GiB6±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB8.23 GiB115.42 GiB3.62 GiB6±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB2.41 GiB115.27 GiB3.77 GiB11±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB1.00 GiB114.08 GiB4.96 GiB12±37%
Step-3.7-FlashIQ4_XS201B99.94 GiB13.30 GiB114.06 GiB4.98 GiB2±12.9%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB6.38 GiB113.87 GiB5.17 GiB6±37%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB11.69 GiB113.77 GiB5.27 GiB2±12.9%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB3.98 GiB113.40 GiB5.64 GiB8±37%
GLM-4.5MoEUD-IQ1_M358B100.32 GiB12.22 GiB113.38 GiB5.66 GiB5±37%
GLM-4.7MoEUD-IQ1_M358B100.27 GiB12.22 GiB113.32 GiB5.72 GiB5±37%
GLM-4.6MoEUD-IQ1_M357B100.02 GiB12.22 GiB113.08 GiB5.96 GiB5±37%
GLM-4.6VMoEQ8_0108B105.81 GiB6.11 GiB112.75 GiB6.29 GiB6±37%
Hy3MoEQ2_K299B101.28 GiB10.63 GiB112.74 GiB6.30 GiB6±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB8.50 GiB112.22 GiB6.82 GiB2±12.9%
MiniMax-M2.7MoEUD-IQ4_NL229B103.15 GiB8.23 GiB112.17 GiB6.87 GiB7±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB3.98 GiB111.81 GiB7.23 GiB8±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_M235B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB6.24 GiB111.80 GiB7.24 GiB6±37%
Trinity-Large-TrueBaseMoEI1-IQ2_S399B108.32 GiB2.41 GiB111.56 GiB7.48 GiB11±37%
MiniMax-M2.1MoEQ3_K_M229B101.77 GiB8.23 GiB110.79 GiB8.25 GiB7±37%
MiniMax-M2.5MoEQ3_K_M229B101.77 GiB8.23 GiB110.79 GiB8.25 GiB7±37%
MiniMax-M2MoEQ3_K_M229B101.77 GiB8.23 GiB110.79 GiB8.25 GiB7±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_M229B101.77 GiB8.23 GiB110.79 GiB8.25 GiB7±37%
GLM-4.7-REAP-218B-A32BMoEQ3_K_M218B97.57 GiB12.22 GiB110.63 GiB8.41 GiB4±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB2.92 GiB110.59 GiB8.45 GiB8±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.14 GiB8.90 GiB13±37%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB7.17 GiB109.86 GiB9.18 GiB2±12.9%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB5.94 GiB109.80 GiB9.24 GiB2±12.9%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB108.98 GiB10.06 GiB13±37%
gemma-2-27b-itF3227.2B101.43 GiB6.54 GiB108.91 GiB10.13 GiB2±12.9%
Devstral-2-123B-Instruct-2512Q6_K125B95.53 GiB11.69 GiB108.18 GiB10.86 GiB2±12.9%
GLM-4.6-Derestricted-v3MoEIQ2_S357B94.86 GiB12.22 GiB107.92 GiB11.12 GiB5±37%
●From the file◐Predictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Questions people ask

What AI models can a NVIDIA DGX Spark (GB10) run?
2091 of 2118 indexed open-weight models fit a NVIDIA DGX Spark (GB10) at 65,536 context with q8_0 KV cache, the largest being Llama-3_3-Nemotron-Super-49B-v1_5 at Q5_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a NVIDIA DGX Spark (GB10) actually have?
Its nameplate is 128 GB, but about 119.04 GiB is available to a model once driver and compositor overhead is accounted for.
Is a NVIDIA DGX Spark (GB10) fast for local AI?
Its memory bandwidth is 273 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.