Lawyer runs DeepSeek-V4.1-Flash on 2x DGX Spark + 3090s: 2,618 tok/s prefill at 256K, 95.9% KLD accuracy, all data open

TumbleweedNew6515 · reddit · 2026-10-31

Nick Mermiges, the lawyer who previously ran local models on a V100 server, is back: DeepSeek-V4.1-Flash-1017 at 2.75 bpw EXL3 on two DGX Sparks with vLLM tensor parallelism, plus an EPYC 3090 box serving Engram tables over RDMA and running the DSpark speculative drafter on NVLinked 3090s.

Repo, launch scripts, bench tool and raw JSON for every cell are public. Everything was driven through Claude Code, which also ran the benchmarks. Upcoming: full 1M-context test, BFCL tool calling, and head-to-head quant comparisons.

Original post →

More from Infra

Infra channel →