Is the DGX Spark faster than the M3 Ultra for LLMs?

The M3 Ultra reads memory about 200% faster than the DGX Spark (819 vs 273 GB/s), so it writes about 200% faster with any model that fits both. In llama.cpp’s measured test the DGX Spark reads a prompt about 2.1× as fast (3,062 vs 1,471 tokens a second). Reading a prompt depends on GPU compute rather than memory bandwidth, so this sets the wait before the first word with a long document or codebase.

Which can run bigger models, the M3 Ultra or the DGX Spark?

The M3 Ultra has about 384 GB for models, the DGX Spark about 96 GB. 60 of 73 popular open models run well on the M3 Ultra, 39 on the DGX Spark.

Can MLX Node run on the DGX Spark?

No. MLX Node and MLX Harness run on Apple Silicon Macs only, with Apple’s MLX framework.

M3 Ultra vs DGX Spark for local LLMs: memory, speed, which to buy

Side by side

	M3 Ultra	DGX Spark
What it is	Mac Studio (2025)	A complete small computer
Memory sizes	96, 256, 512 GB	128 GB
Memory bandwidth	819 GB/s	273 GB/s
For models, at most	about 384 GB	about 96 GB
Models that run well	60 of 73	39 of 73
Measured: writing	92 tok/s	57 tok/s
Measured: reading a prompt	1,471 tok/s	3,062 tok/s

Measured rows: each machine’s fastest result in llama.cpp’s public benchmarks with Llama 2 7B Q4_0, writing 128 tokens (tg128) and reading a 512-token prompt (pp512), in tokens a second. The Mac ran Metal, the DGX Spark CUDA. Apple Silicon results · DGX Spark results For models we count about 75% of memory on both, the macOS default, because the DGX Spark also shares one memory pool between its CPU and GPU. How much its GPU may really use depends on its operating system and settings.

Where each one wins

Writing speed. The M3 Ultra reads memory about 200% faster than the DGX Spark (819 vs 273 GB/s), so it writes about 200% faster with any model that fits both.
Model size. The M3 Ultra has about 384 GB for models, the DGX Spark about 96 GB. 60 of 73 popular open models run well on the M3 Ultra, 39 on the DGX Spark.
Reading long prompts. In llama.cpp’s measured test the DGX Spark reads a prompt about 2.1× as fast (3,062 vs 1,471 tokens a second). Reading a prompt depends on GPU compute rather than memory bandwidth, so this sets the wait before the first word with a long document or codebase.
Software. The DGX Spark runs NVIDIA’s CUDA, which most AI software supports first, including training code and many image and video tools. On a Mac you run models with MLX, llama.cpp’s Metal backend or PyTorch’s MPS backend instead.
Around it. Both are complete computers that keep the model in one memory pool shared by the CPU and GPU.
The Mac. It’s also an everyday computer, and it runs MLX Harness and MLX Node, which are made for Apple Silicon.

Speed with popular models

Tokens a second for one conversation, each chip at its largest memory size (512 GB and 128 GB), with the best version of the model that fits.

Model	M3 Ultra	DGX Spark
Llama 3.1 8B Instruct8B	43–77 8-bit	14–26 8-bit
gpt-oss-20b21B · 3.6B active	194–345 native 4.25-bit	65–115 native 4.25-bit
Qwen3.6 35B A3B35B · 3B active	115–205 8-bit	38–68 8-bit
Qwen3.8 27B27B	13–23 8-bit	4–8 8-bit
Llama 3.3 70B Instruct70B	5–9 8-bit	2–3 8-bit
gpt-oss-120b117B · 5.1B active	137–243 native 4.25-bit	45–81 native 4.25-bit
Qwen3.5 122B A10B122B · 10B active	35–62 8-bit	22–39 4-bit
DeepSeek V4 Flash 0731284B · 13B active	27–47 8-bit	30–53 2-bit lower quality

Questions

Is the DGX Spark faster than the M3 Ultra for LLMs?: The M3 Ultra reads memory about 200% faster than the DGX Spark (819 vs 273 GB/s), so it writes about 200% faster with any model that fits both. In llama.cpp’s measured test the DGX Spark reads a prompt about 2.1× as fast (3,062 vs 1,471 tokens a second). Reading a prompt depends on GPU compute rather than memory bandwidth, so this sets the wait before the first word with a long document or codebase.
Which can run bigger models, the M3 Ultra or the DGX Spark?: The M3 Ultra has about 384 GB for models, the DGX Spark about 96 GB. 60 of 73 popular open models run well on the M3 Ultra, 39 on the DGX Spark.
Can MLX Node run on the DGX Spark?: No. MLX Node and MLX Harness run on Apple Silicon Macs only, with Apple’s MLX framework.

More comparisons

The M3 Ultra for local AI · Compare every model · The best Mac for local LLMs

How it’s worked out

Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.

DGX Spark figures: NVIDIA: DGX Spark (specs).