What is the biggest model the NVIDIA DGX Spark can run?

gpt-oss 120B, which needs about 67 GB, on the 128 GB version at about 33 tok/s.

What is the biggest model the Jetson AGX Orin can run?

Llama 3.3 70B, which needs about 45 GB, on the 64 GB version at about 3 tok/s.

NVIDIA DGX Spark vs Jetson AGX Orin for local AI

The NVIDIA DGX Spark fits more: up to gpt-oss 120B, against Llama 3.3 70B on the Jetson AGX Orin. The NVIDIA DGX Spark generates about 33% faster on any model both can run, because it has 273 GB/s of memory bandwidth to the Jetson AGX Orin's 205 GB/s.

The numbers that matter

Memory and bandwidth

Memory decides which models fit. Bandwidth decides how fast they generate, because every token reads the model's active weights once.

	NVIDIA DGX Spark	Jetson AGX Orin
Memory	128 GB unified memory	32 or 64 GB unified memory
Usable for models	About 96 GB of 128 GB	About 48 GB of 64 GB
Bandwidth	273 GB/s	205 GB/s
Biggest model that fits	gpt-oss 120B	Llama 3.3 70B
Best everyday pick	gpt-oss 120B, ~33 tok/s	Qwen3 30B-A3B, ~36 tok/s

Model by model

Every model, side by side

Estimated generation speed at the largest memory size each was sold in (128 GB and 64 GB).

Model	NVIDIA DGX Spark	Jetson AGX Orin
Qwen3 4B About 4 GB	~60 tok/s Runs great	~45 tok/s Runs great
Qwen3 8B About 7 GB	~29 tok/s Usable	~22 tok/s Usable
Gemma 3 12B About 10 GB	~19 tok/s Usable	~14 tok/s Usable
Qwen3 14B About 11 GB	~16 tok/s Usable	~12 tok/s Usable
gpt-oss 20B About 16 GB	~48 tok/s Runs great	~36 tok/s Runs great
Gemma 3 27B About 19 GB	~9 tok/s Slow	~7 tok/s Slow
Qwen3 30B-A3B About 21 GB	~48 tok/s Runs great	~36 tok/s Runs great
Qwen3 32B About 22 GB	~8 tok/s Slow	~6 tok/s Slow
Llama 3.3 70B About 45 GB	~3 tok/s Slow	~3 tok/s Slow
gpt-oss 120B About 67 GB	~33 tok/s Runs great	Needs 67 GB
Qwen3 235B-A22B About 144 GB	Needs 144 GB	Needs 144 GB
DeepSeek R1 671B About 406 GB	Needs 406 GB	Needs 406 GB

Which should you get?

The NVIDIA DGX Spark: it fits bigger models and runs faster. Price matters as much as either number: check used prices on the builds page, where every parts list has a measured or estimated speed.

How we estimate

Both columns use the same formula: memory bandwidth divided by the weights read per token, calibrated on our own rack, with Ollama's default 4-bit sizes. Real results are usually within about a third. Measured either one? Send us the number.

Compare

Other matchups

Every comparison