Community
DeepSeek-V4.1-Flash runs on two DGX Sparks with a 3090 offload box, reporting high prefill speeds
A community builder reports 2,618 to 2,827 tok/s prefill up to 256K on a 2.75 bpw EXL3 quant using vLLM across two DGX Sparks, with an EPYC box holding Engram tables and running a speculative drafter. Decode is slower than some rival builds, and the author says the benchmarks are self-run.