DGX Spark (GB10, 128 GB) vs M3 Ultra (512 GB, 80-core GPU)

The DGX Spark (GB10, 128 GB) runs 103 of 110 open models at 4-bit or better, the M3 Ultra (512 GB, 80-core GPU) runs 109. Usable memory: about 122 GB against 384 GB. The M3 Ultra (512 GB, 80-core GPU) has more memory bandwidth (819 vs 273 GB/s), so models that fit both answer faster on it.

Data as of 2 November 2026. Speeds are estimates unless marked measured.

NVIDIA

DGX Spark

128 GB

All models for this machine

Apple

M3 Ultra

512 GB · 80-core GPU

All models for this machine

Running gpt-oss-120b

Runs

~61 tok/s

MXFP4 · 68 of 122 GB · reported

Runs

~101 tok/s

MXFP4 · 68 of 384 GB · ModelFit est.

Single-user decode speed, estimated. Fastest machines for this model

Memory

128 GB

memory

512 GB

memory

~122 GB

usable for a model

~384 GB

usable for a model

273 GB/s

memory bandwidth

819 GB/s

memory bandwidth

What it runs

103

of 110 models run at 4-bit or better

109

of 110 models run at 4-bit or better

Our pick

Qwen3.8-Flash-Next

Q4_K_XL · ~28 tok/s our est.

GLM-5.3-Flash

Q8_0 · ~24 tok/s our est.

Largest current model that fits

Qwen3.8-Flash-Next

125B at Q4_K_XL

GLM-5.3-Flash

320B at Q8_0

Popular models

gpt-oss-120b

Runs

MXFP4 · ~61 tok/s

Runs

MXFP4 · ~101 tok/s

gpt-oss-20b

Runs

MXFP4 · ~63 tok/s

Runs

MXFP4 · ~117 tok/s

Qwen3.6-35B-A3B

Runs

Q8_0 · ~45 tok/s

Runs

Q8_0 · ~112 tok/s

Qwen3.8-27B

Runs

Q4_K_M · ~11 tok/s

Runs

Q8_0 · ~22 tok/s

Gemma 4 26B-A4B

Runs

Q8_0 · ~52 tok/s

Runs

Q8_0 · ~85 tok/s

Llama 3.3 70B

Runs

Q4_K_M · ~6.0 tok/s slow

Runs

Q6_K · ~11 tok/s

Nemotron 3 Super 120B-A12B

Runs

Q4_K_M · ~23 tok/s

Runs

Q4_K_M · ~41 tok/s

DeepSeek-V4-Flash

2-bit only

Q2_K · ~13 tok/s

Runs

Q8_0 · ~33 tok/s

Rent one by the hour

Use Change under a machine to pick another type, producer, model or variant. How the numbers are worked out