Running DeepSeek V4.1 Flash 1017 on a single DGX Spark (ds4-dfm-rs, Docker Compose)

This GitHub repository provides a complete solution for deploying the quantized version of DeepSeek V4.1 Flash 1017 on a single NVIDIA DGX Spark with one command; the routed experts are vector-quantized at about 1.6 bit and the rest is q4_K, the 114 GB base file resides in memory, and the 203 GB Engram lookup table stays on the SSD and is read token by token. The first run requires building the engine and downloading 318 GB of files, which the author says takes about 2.5 hours, after which startup takes about 2 minutes; the engine ds4-dfm-rs v0.1.3 serves only one request at a time, the context is 1M tokens, and the model card states that this quantization selects the same next token as the full model in 69-79% of cases. The author recommends that for tools involving transfers or sending messages, a payee whitelist be set outside the model or a confirmation step be added for newly added recipients.