Can the NVIDIA DGX Spark run Llama 3.3 70B?
Yes. Llama 3.3 70B needs about 45 GB and the NVIDIA DGX Spark with 128 GB of unified memory leaves about 96 GB for models. Expect about 3 tok/s, slow, but it works.
ollama run llama3.3:70bFirst time? Install Ollama, then paste that into a terminal.
Llama 3.3 70B needsAbout 45 GB
NVIDIA DGX Spark has128 GB unified memory
Bandwidth273 GB/s
Estimated speed~3 tok/s
Want it faster?
Speed comes from memory bandwidth: the card reads all 43 GB of active weights for every token. On the NVIDIA DGX Spark, gpt-oss 120B runs at about 33 tok/s if you can trade some quality for speed. Cards that run Llama 3.3 70B faster:
How we estimate
Speed is roughly memory bandwidth divided by the weights read per token, calibrated on our own rack. Sizes are Ollama's default 4-bit downloads plus about 1.2 GB for context. Real results are usually within about a third. Measured it yourself? Send us the number.
Keep going