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 4K context with f16 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 4K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2-FlashMoEKV unresolvedIQ3_XS310B117.68 GiB0.47 GiB119.00 GiB0.04 GiB10±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.73 GiB118.80 GiB0.24 GiB8±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.73 GiB118.80 GiB0.24 GiB8±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.12 GiB118.78 GiB0.26 GiB12±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.50 GiB118.72 GiB0.32 GiB8±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.09 GiB118.71 GiB0.33 GiB11±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.47 GiB118.59 GiB0.45 GiB10±37%
step-3.5-flashQ4_1199B115.15 GiB2.53 GiB118.51 GiB0.53 GiB2±12.9%
GLM-4.7-REAP-218B-A32BMoEIQ4_NL218B116.20 GiB1.44 GiB118.47 GiB0.57 GiB6±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.73 GiB118.45 GiB0.59 GiB8±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.84 GiB118.33 GiB0.71 GiB2±12.9%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.94 GiB118.27 GiB0.77 GiB12±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.94 GiB118.27 GiB0.77 GiB12±37%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.73 GiB118.26 GiB0.78 GiB8±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.73 GiB118.26 GiB0.78 GiB8±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ4_XS235B116.68 GiB0.73 GiB118.25 GiB0.79 GiB8±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.26 GiB118.04 GiB1.00 GiB10±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.26 GiB118.04 GiB1.00 GiB10±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.26 GiB118.04 GiB1.00 GiB10±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.12 GiB117.39 GiB1.65 GiB11±37%
grok-2MoEIQ3_M270B115.25 GiB1.00 GiB117.19 GiB1.85 GiB3±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB2.53 GiB117.06 GiB1.98 GiB2±12.9%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.12 GiB116.80 GiB2.24 GiB12±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.17 GiB116.72 GiB2.32 GiB11±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB1.44 GiB116.31 GiB2.73 GiB8±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB1.44 GiB116.31 GiB2.73 GiB8±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB1.44 GiB115.84 GiB3.20 GiB8±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB0.97 GiB115.75 GiB3.29 GiB10±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB0.97 GiB115.53 GiB3.51 GiB10±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB0.97 GiB115.53 GiB3.51 GiB10±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB0.97 GiB115.28 GiB3.76 GiB10±37%
Hermes-4-405BIQ2_XS406B111.15 GiB1.97 GiB114.20 GiB4.84 GiB2±12.9%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB1.97 GiB114.20 GiB4.84 GiB2±12.9%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.75 GiB114.05 GiB4.99 GiB14±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.94 GiB113.79 GiB5.25 GiB12±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.75 GiB113.21 GiB5.83 GiB11±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.12 GiB113.20 GiB5.84 GiB13±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB0.97 GiB111.98 GiB7.06 GiB10±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.72 GiB110.94 GiB8.10 GiB8±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.72 GiB110.94 GiB8.10 GiB8±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB1.44 GiB110.75 GiB8.29 GiB8±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.06 GiB110.16 GiB8.88 GiB13±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.47 GiB109.89 GiB9.15 GiB11±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB0.97 GiB109.65 GiB9.39 GiB8±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB1.44 GiB109.42 GiB9.62 GiB9±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.88 GiB109.33 GiB9.71 GiB3±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.06 GiB109.01 GiB10.03 GiB13±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.75 GiB108.24 GiB10.80 GiB8±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB0.97 GiB108.16 GiB10.88 GiB9±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB0.97 GiB108.16 GiB10.88 GiB9±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.34 GiB108.01 GiB11.03 GiB10±37%
GLM-4.6VMoEQ8_0108B105.81 GiB0.72 GiB107.36 GiB11.68 GiB9±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB10.00 GiB106.88 GiB12.16 GiB2±12.9%
dots.llm1.instMoEQ5_K_M143B101.84 GiB3.88 GiB106.55 GiB12.49 GiB8±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.12 GiB106.36 GiB12.68 GiB11±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.80 GiB105.66 GiB13.38 GiB2±12.9%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB1.00 GiB104.72 GiB14.32 GiB2±12.9%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB10.00 GiB103.84 GiB15.20 GiB2±12.9%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB10.00 GiB103.84 GiB15.20 GiB2±12.9%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB10.00 GiB103.84 GiB15.20 GiB2±12.9%
●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 4,096 context with f16 KV cache, the largest being MiMo-V2-Flash at IQ3_XS. 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.