Uploaded September 2026 | Updated November 2026, 3 hours ago
Two NVIDIA DGX Sparks, one model. I ran DeepSeek-V4-Flash, 284B params, full 1M-token context, across two DGX Sparks linked by QSFP/RoCE, driven entirely from my Mac over SSH.
The 30-megabyte mistake that froze both nodes, the off-by-one that wasn't, and every script I wrote so it never happens twice.
Full story: dashaun.com/posts/deepseek-v4-flash-2x-dgx-spark
Repo: github.com/dashaun/deepseek-v4-flash-2x-dgx-spark
Two NVIDIA DGX Sparks, one model. I ran DeepSeek-V4-Flash, 284B params, full 1M-token context, across two DGX Sparks linked by QSFP/RoCE, driven entirely from my Mac over SSH.
The 30-megabyte mistake that froze both nodes, the off-by-one that wasn't, and every script I wrote so it never happens twice.
Full story: dashaun.com/posts/deepseek-v4-flash-2x-dgx-spark
Repo: github.com/dashaun/deepseek-v4-flash-2x-dgx-spark










