StudioLawyer runs DeepSeek-V4.1-Flash on 2x DGX Spark plus a 3090 box, hitting 2,600+ tok/s prefill to 256K and publishing the full stack
Original title: Lawyer with too many GPUs, part 3: DeepSeek-V4.1-Flash on 2x DGX Spark + an old 3090 box. 2,600+ tok/s prefill to 256K, 322 tok/s structured output, full stack and raw data published
The author runs DeepSeek-V4.1-Flash-1017 at 2.75 bpw EXL3 with vLLM tensor parallelism across two DGX Sparks, while an EPYC box with 3090s serves Engram tables over RDMA and runs the DSpark speculative drafter. Prefill reaches 2,618 tok/s at 256K, structured output hits 322 tok/s at 4 streams (138 single), and the ~1.93M-token KV cache plus NVMe prefix cache cuts first token after a 150K prompt restart from 58.6 s to 2.58 s. Code, launch scripts, the bench tool and raw JSON are all public at https://github.com/ProprietaryLegal/DeepSeek-V4.1-Flash-2xDGX-Spark-x86-offload (prefix-cache code pending); the author concedes slower plain prose decoding and weaker long-context decode than MiaAI's and bertholomus's builds.
The author publishes the full deployment stack and raw benchmark data for a 2x DGX Spark plus 3090 offload build, letting readers compare it item by item with similar setups.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.