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 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 4K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-235B-A22B-abliteratedMoEIQ4_XS235B117.97 GiB0.21 GiB119.01 GiB0.03 GiB8±37%
GLM-4.7-REAP-218B-A32BMoEQ4_0218B117.58 GiB0.40 GiB118.82 GiB0.22 GiB7±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.03 GiB118.69 GiB0.35 GiB12±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XS310B117.68 GiB0.13 GiB118.66 GiB0.38 GiB11±37%
Hermes-4-405BIQ2_S406B117.02 GiB0.55 GiB118.65 GiB0.39 GiB2±12.9%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.02 GiB118.64 GiB0.40 GiB11±37%
Hy3MoEIQ3_XXS299B117.43 GiB0.35 GiB118.62 GiB0.42 GiB9±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.14 GiB118.36 GiB0.68 GiB8±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.21 GiB118.28 GiB0.76 GiB8±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.21 GiB118.28 GiB0.76 GiB8±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.13 GiB118.25 GiB0.79 GiB11±37%
Step-3.7-FlashQ4_1201B116.67 GiB0.71 GiB118.21 GiB0.83 GiB2±12.9%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.21 GiB117.92 GiB1.12 GiB8±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.07 GiB117.85 GiB1.19 GiB10±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.07 GiB117.85 GiB1.19 GiB10±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.07 GiB117.85 GiB1.19 GiB10±37%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.21 GiB117.74 GiB1.30 GiB8±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.21 GiB117.74 GiB1.30 GiB8±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.24 GiB117.72 GiB1.32 GiB2±12.9%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.26 GiB117.59 GiB1.45 GiB13±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.26 GiB117.59 GiB1.45 GiB13±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.03 GiB117.30 GiB1.74 GiB11±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.03 GiB116.71 GiB2.33 GiB12±37%
step-3.5-flashQ4_1199B115.15 GiB0.71 GiB116.69 GiB2.35 GiB2±12.9%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.05 GiB116.60 GiB2.44 GiB11±37%
grok-2MoEIQ3_M270B115.25 GiB0.28 GiB116.47 GiB2.57 GiB3±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB0.40 GiB115.28 GiB3.76 GiB9±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB0.40 GiB115.28 GiB3.76 GiB9±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB0.27 GiB115.05 GiB3.99 GiB11±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB0.27 GiB114.83 GiB4.21 GiB11±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB0.27 GiB114.83 GiB4.21 GiB11±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB0.40 GiB114.80 GiB4.24 GiB9±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB0.27 GiB114.59 GiB4.45 GiB11±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.21 GiB113.51 GiB5.53 GiB15±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.26 GiB113.12 GiB5.92 GiB13±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.03 GiB113.12 GiB5.92 GiB13±37%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB0.55 GiB112.79 GiB6.25 GiB2±12.9%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.21 GiB112.67 GiB6.37 GiB12±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB0.27 GiB111.28 GiB7.76 GiB11±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.20 GiB110.42 GiB8.62 GiB9±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.20 GiB110.42 GiB8.62 GiB9±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.12 GiB8.92 GiB13±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB0.40 GiB109.72 GiB9.32 GiB8±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.13 GiB109.55 GiB9.49 GiB12±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB108.96 GiB10.08 GiB13±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB0.27 GiB108.96 GiB10.08 GiB8±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.25 GiB108.70 GiB10.34 GiB4±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB0.40 GiB108.39 GiB10.65 GiB10±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.10 GiB107.76 GiB11.28 GiB10±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.21 GiB107.70 GiB11.34 GiB9±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB0.27 GiB107.46 GiB11.58 GiB10±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB0.27 GiB107.46 GiB11.58 GiB10±37%
GLM-4.6VMoEQ8_0108B105.81 GiB0.20 GiB106.84 GiB12.20 GiB9±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.03 GiB106.28 GiB12.76 GiB11±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB0.51 GiB104.37 GiB14.67 GiB2±12.9%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB0.28 GiB104.00 GiB15.04 GiB2±12.9%
dots.llm1.instMoEQ5_K_M143B101.84 GiB1.09 GiB103.76 GiB15.28 GiB10±37%
gpt-oss-20b-hereticMoEQ8_020.9B102.72 GiB0.03 GiB103.54 GiB15.50 GiB6±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ8_020.9B102.72 GiB0.03 GiB103.54 GiB15.50 GiB6±37%
gemma-2-27b-itF3227.2B101.43 GiB0.40 GiB102.77 GiB16.27 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 q4_0 KV cache, the largest being Qwen3-235B-A22B-abliterated at IQ4_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.