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 128K context with q4_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 128K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
command-a-plus-05-2026-bf16MoEIQ4_NL219B116.79 GiB1.24 GiB118.84 GiB0.20 GiB8±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB17.72 GiB118.71 GiB0.33 GiB2±12.9%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB17.72 GiB118.71 GiB0.33 GiB2±12.9%
Qwen3.5-397B-A17BMoEIQ2_XS403B116.79 GiB1.05 GiB118.70 GiB0.34 GiB11±37%
step-3.5-flashIQ4_NL199B103.97 GiB13.79 GiB118.59 GiB0.45 GiB2±12.9%
dots.llm1.instMoEQ4_K_S143B82.81 GiB34.88 GiB118.52 GiB0.52 GiB3±37%
Llama-3_1-Nemotron-51B-InstructQ4_K_S51.5B27.46 GiB90.00 GiB118.40 GiB0.64 GiB2±12.9%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB1.05 GiB118.32 GiB0.72 GiB10±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB4.22 GiB118.21 GiB0.83 GiB8±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB1.27 GiB117.83 GiB1.21 GiB10±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB1.05 GiB117.74 GiB1.30 GiB11±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q4_K_S49.9B26.67 GiB90.00 GiB117.61 GiB1.43 GiB2±12.9%
Valkyrie-49B-v2.1I1-Q4_K_S49.9B26.67 GiB90.00 GiB117.61 GiB1.43 GiB2±12.9%
Llama-3_3-Nemotron-Super-49B-v1Q4_K_S49.9B26.67 GiB90.00 GiB117.61 GiB1.43 GiB2±12.9%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB8.72 GiB117.40 GiB1.64 GiB5±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ3_K_L236B113.97 GiB2.37 GiB117.18 GiB1.86 GiB9±37%
DeepSeek-V2.5MoEQ3_K_L236B113.97 GiB2.37 GiB117.18 GiB1.86 GiB9±37%
DeepSeek-Coder-V2-InstructMoEQ3_K_L236B113.97 GiB2.37 GiB117.18 GiB1.86 GiB9±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB6.47 GiB116.69 GiB2.35 GiB6±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB6.47 GiB116.69 GiB2.35 GiB6±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ3_XXS269B102.71 GiB12.94 GiB116.48 GiB2.56 GiB4±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB7.88 GiB116.33 GiB2.71 GiB3±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB8.72 GiB115.91 GiB3.13 GiB6±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB8.72 GiB115.91 GiB3.13 GiB6±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB2.33 GiB115.19 GiB3.85 GiB11±37%
Step-3.7-FlashIQ4_XS201B99.94 GiB13.79 GiB114.55 GiB4.49 GiB2±12.9%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB12.38 GiB114.46 GiB4.58 GiB2±12.9%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB6.75 GiB114.24 GiB4.80 GiB6±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB1.05 GiB114.14 GiB4.90 GiB12±37%
GLM-4.5MoEUD-IQ1_M358B100.32 GiB12.94 GiB114.10 GiB4.94 GiB5±37%
grok-2MoEIQ3_XS270B104.12 GiB9.00 GiB114.07 GiB4.97 GiB3±37%
GLM-4.7MoEUD-IQ1_M358B100.27 GiB12.94 GiB114.04 GiB5.00 GiB5±37%
GLM-4.6MoEUD-IQ1_M357B100.02 GiB12.94 GiB113.80 GiB5.24 GiB5±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB4.22 GiB113.64 GiB5.40 GiB8±37%
Hy3MoEQ2_K299B101.28 GiB11.25 GiB113.37 GiB5.67 GiB5±37%
GLM-4.6VMoEQ8_0108B105.81 GiB6.47 GiB113.11 GiB5.93 GiB6±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB9.00 GiB112.72 GiB6.32 GiB2±12.9%
MiniMax-M2.7MoEUD-IQ4_NL229B103.15 GiB8.72 GiB112.65 GiB6.39 GiB6±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_M235B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB6.61 GiB112.17 GiB6.87 GiB6±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB4.22 GiB112.04 GiB7.00 GiB8±37%
Trinity-Large-TrueBaseMoEI1-IQ2_S399B108.32 GiB2.33 GiB111.48 GiB7.56 GiB11±37%
GLM-4.7-REAP-218B-A32BMoEQ3_K_M218B97.57 GiB12.94 GiB111.35 GiB7.69 GiB4±37%
MiniMax-M2.1MoEQ3_K_M229B101.77 GiB8.72 GiB111.27 GiB7.77 GiB6±37%
MiniMax-M2.5MoEQ3_K_M229B101.77 GiB8.72 GiB111.27 GiB7.77 GiB6±37%
MiniMax-M2MoEQ3_K_M229B101.77 GiB8.72 GiB111.27 GiB7.77 GiB6±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_M229B101.77 GiB8.72 GiB111.27 GiB7.77 GiB6±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB3.09 GiB110.76 GiB8.28 GiB8±37%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB7.59 GiB110.28 GiB8.76 GiB2±12.9%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.12 GiB8.92 GiB13±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB5.95 GiB109.82 GiB9.22 GiB2±12.9%
gemma-2-27b-itF3227.2B101.43 GiB6.70 GiB109.06 GiB9.98 GiB2±12.9%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB108.96 GiB10.08 GiB13±37%
Devstral-2-123B-Instruct-2512Q6_K125B95.53 GiB12.38 GiB108.86 GiB10.18 GiB2±12.9%
GLM-4.6-Derestricted-v3MoEIQ2_S357B94.86 GiB12.94 GiB108.64 GiB10.40 GiB5±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB1.05 GiB107.30 GiB11.74 GiB10±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 131,072 context with q4_0 KV cache, the largest being command-a-plus-05-2026-bf16 at IQ4_NL. 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.