NVIDIA DGX Spark 128GB vs Ryzen AI Max+ 395 128GB for Local LLMs

The NVIDIA DGX Spark 128GB holds bigger models (112 GB vs 96.0 GB), and the NVIDIA DGX Spark 128GB generates tokens about 7% faster. Below: speeds on seven popular models, the models only one card can run, and our verdict.

Updated October 14, 2026 · Estimates from memory bandwidth, see methodology

Specs that matter for AI

NVIDIA DGX Spark 128GBRyzen AI Max+ 395 128GB
Memory128 GB unified128 GB unified
GPU-usable memory112 GB96.0 GB
Bandwidth273 GB/s256 GB/s
ArchitectureGrace BlackwellStrix Halo
Released2025-102025-02
Llama 3.1 8B Q4_K_M~30 tok/s~28 tok/s

Speed and fit by model

Q4_K_M (or native format) with 8K context.

ModelVRAMNVIDIA DGX Spark 128GBRyzen AI Max+ 395 128GB
Llama 3.1 8B6.3 GBFits ~30 tok/sFits ~28 tok/s
Qwen 3 14B10.5 GBFits ~17 tok/sFits ~16 tok/s
gpt-oss-20b13.9 GBFits ~57 tok/sFits ~54 tok/s
Qwen 3 30B-A3B18.8 GBFits ~51 tok/sFits ~49 tok/s
Qwen 3 32B21.4 GBFits ~7.8 tok/sFits ~7.3 tok/s
Llama 3.3 70B43.2 GBFits ~3.7 tok/sFits ~3.5 tok/s
gpt-oss-120b62.0 GBFits ~46 tok/sFits ~43 tok/s

Verdict

Both are 128 GB unified-memory boxes with similar bandwidth (273 vs 256 GB/s), so token generation on the same model is close. The DGX Spark lets more memory go to the GPU and runs the full CUDA stack, which matters for training and vLLM. The Ryzen AI Max systems are cheaper and double as a normal PC. Both suit MoE models like gpt-oss-120b far better than large dense models.

Which should you buy?

Choose the NVIDIA DGX Spark 128GB if…

  • You develop with CUDA, vLLM or PyTorch.
  • You want to fine-tune small models locally.
  • You want the most GPU-usable memory in a small box.

Choose the Ryzen AI Max+ 395 128GB if…

  • You want the lower price.
  • You also want a normal Windows or Linux PC.
  • You mainly run llama.cpp-based apps.

Best quality and longest context on each card

The highest-quality quantization that fits fully at 8K context, and the longest context that fits at Q4_K_M with an FP16 cache.

ModelNVIDIA DGX Spark 128GB: best quantRyzen AI Max+ 395 128GB: best quantNVIDIA DGX Spark 128GB: max contextRyzen AI Max+ 395 128GB: max context
Qwen 3.5 9BFP16 / BF16FP16 / BF16256K256K
Qwen 3 14BFP16 / BF16FP16 / BF1640K40K
gpt-oss-20bMXFP4 (official)MXFP4 (official)128K128K
Qwen 3.6 27BFP16 / BF16FP16 / BF16256K256K
Qwen 3 32BFP16 / BF16FP16 / BF1640K40K
Llama 3.3 70BQ8_0Q8_0128K128K

Image generation

Best weight precision for image models with model swapping, at each model's default resolution. Image generation is also limited by compute, so the card with more tensor-core throughput finishes each image faster even when both fit.

ModelNVIDIA DGX Spark 128GBRyzen AI Max+ 395 128GB
SDXL 1.0FP16 / BF16FP16 / BF16
FLUX.1 [dev]FP16 / BF16FP16 / BF16
Qwen-ImageFP16 / BF16FP16 / BF16
Wan 2.2 A14B (T2V / I2V)FP16 / BF16FP16 / BF16

Models only one card can run

Only NVIDIA DGX Spark 128GB

Only Ryzen AI Max+ 395 128GB

None

Frequently asked questions

Is the NVIDIA DGX Spark 128GB or Ryzen AI Max+ 395 128GB better for local LLMs?

Both are 128 GB unified-memory boxes with similar bandwidth (273 vs 256 GB/s), so token generation on the same model is close. The DGX Spark lets more memory go to the GPU and runs the full CUDA stack, which matters for training and vLLM. The Ryzen AI Max systems are cheaper and double as a normal PC. Both suit MoE models like gpt-oss-120b far better than large dense models.

Which is faster for LLM inference?

The NVIDIA DGX Spark 128GB, by about 7% for token generation on models that fit both (about 30 vs 28 tokens/s on Llama 3.1 8B Q4_K_M).

Which can run bigger models?

The NVIDIA DGX Spark 128GB, with 112 GB usable versus 96.0 GB.

Related