DeepSeek V4.1 Flash: Strix Halo SSD vs cluster setup

This AMD-sponsored tutorial compares two ways to run DeepSeek V4.1 Flash through DwarfStar on Ryzen AI MAX+ 395 hardware. The presenter uses antirez's calibrated 2-bit quantization: about 366GB in total, including roughly 203GB of Engram tables that stay on disk. The remaining 163GB exceeds one 128GB host's memory, so that setup needs SSD streaming. Two 128GB machines can keep the main-model weights in memory.

The architecture explanation covers the 552-billion-parameter mixture-of-experts backbone, causal encoder-decoder design and Compressed Sparse Attention v2. The presenter explains that the encoder reduces long-prompt processing work; generated tokens still pass through all 40 layers.

Setup uses a prebuilt container in AI Toolbox Cockpit. A cluster needs worker and coordinator roles, the coordinator IP in both configurations, and model files on both hosts. Support was included in the presenter's toolbox but had not merged upstream at recording time.

For a fresh 65,000-token prompt, the presenter reports about 351 tokens/s prefill and 8.5 tokens/s generation on one host, versus 430 and 14 to 15 on two. These are early implementation results. Terminal Bench Mini completed all 19 tasks, one on a second attempt; the presenter treats this as an implementation and quantization check rather than a model ranking. DSpark support remained future work.