DeepSeek v4 Flash on Two DGX Sparks - realistic tok/sec?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

I am considering buy 2 DGX Sparks for a total of 256gb of unified RAM, just for running models like deepseek-v4-flash.

I saw recent posts here, but I can’t seems to find a real number for my use case.

For a big context, like 256k, how much tok/sec would I get in this setup?

Thanks in advance

2

Take a look at this thread – several benchmarks and configs that could serve as basis for you: DeepSeek-V4-Flash (official FP8) running across 2x DGX Spark — TP=2, MTP, 200K ctx, recipe + numbers - #53 by serapis

3

I am getting around 25