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 16K 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 16K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_XS310B117.12 GiB1.00 GiB118.96 GiB0.08 GiB10±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.25 GiB118.91 GiB0.13 GiB12±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.19 GiB118.81 GiB0.23 GiB11±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.49 GiB118.71 GiB0.33 GiB8±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.56 GiB118.34 GiB0.70 GiB10±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.56 GiB118.34 GiB0.70 GiB10±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.56 GiB118.34 GiB0.70 GiB10±37%
grok-2MoEIQ3_M270B115.25 GiB2.13 GiB118.32 GiB0.72 GiB3±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB3.74 GiB118.27 GiB0.77 GiB2±12.9%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.92 GiB118.25 GiB0.79 GiB12±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.92 GiB118.25 GiB0.79 GiB12±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB21.25 GiB118.13 GiB0.91 GiB2±12.9%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB3.05 GiB117.93 GiB1.11 GiB7±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB3.05 GiB117.93 GiB1.11 GiB7±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.25 GiB117.52 GiB1.52 GiB11±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB3.05 GiB117.45 GiB1.59 GiB7±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.25 GiB116.93 GiB2.11 GiB12±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.31 GiB116.87 GiB2.17 GiB11±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB2.06 GiB116.84 GiB2.20 GiB9±37%
step-3.5-flashQ4_K_L199B112.07 GiB3.74 GiB116.63 GiB2.41 GiB2±12.9%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB2.06 GiB116.62 GiB2.42 GiB9±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB2.06 GiB116.62 GiB2.42 GiB9±37%
Hermes-4-405BIQ2_XS406B111.15 GiB4.18 GiB116.42 GiB2.62 GiB2±12.9%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB4.18 GiB116.42 GiB2.62 GiB2±12.9%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB2.06 GiB116.37 GiB2.67 GiB9±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB1.56 GiB115.85 GiB3.19 GiB8±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB21.25 GiB115.09 GiB3.95 GiB2±12.9%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB21.25 GiB115.09 GiB3.95 GiB2±12.9%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB21.25 GiB115.09 GiB3.95 GiB2±12.9%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB1.59 GiB114.90 GiB4.14 GiB12±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB1.59 GiB114.05 GiB4.99 GiB10±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB3.05 GiB113.99 GiB5.05 GiB6±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.92 GiB113.78 GiB5.26 GiB12±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.25 GiB113.33 GiB5.71 GiB13±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB2.06 GiB113.07 GiB5.97 GiB10±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB3.05 GiB112.37 GiB6.67 GiB7±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB1.53 GiB111.74 GiB7.30 GiB8±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB1.53 GiB111.74 GiB7.30 GiB8±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB3.05 GiB111.04 GiB8.00 GiB8±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB8.23 GiB110.91 GiB8.13 GiB6±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB2.06 GiB110.74 GiB8.30 GiB7±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB1.00 GiB110.41 GiB8.63 GiB11±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB1.86 GiB110.32 GiB8.72 GiB3±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.14 GiB8.90 GiB13±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB2.06 GiB109.25 GiB9.79 GiB8±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB2.06 GiB109.25 GiB9.79 GiB8±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB1.59 GiB109.09 GiB9.95 GiB8±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB108.98 GiB10.06 GiB13±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB1.00 GiB108.82 GiB10.22 GiB11±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.73 GiB108.40 GiB10.64 GiB10±37%
GLM-4.6VMoEQ8_0108B105.81 GiB1.53 GiB108.16 GiB10.88 GiB8±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB1.56 GiB107.12 GiB11.92 GiB8±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB1.56 GiB107.12 GiB11.92 GiB8±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB1.56 GiB107.12 GiB11.92 GiB8±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB1.56 GiB107.12 GiB11.92 GiB8±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB1.56 GiB107.12 GiB11.92 GiB8±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.25 GiB106.49 GiB12.55 GiB11±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB2.13 GiB105.84 GiB13.20 GiB2±12.9%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.95 GiB105.81 GiB13.23 GiB2±12.9%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB2.92 GiB105.01 GiB14.03 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 16,384 context with q8_0 KV cache, the largest being MiMo-V2-Flash at I1-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.
NVIDIA DGX Spark (GB10) — what AI models can it run locally? — ossmodeldb