elevatormusic/deepseek-v4.1-flash-1017-exl3-tensorfold-2x-dgx-spark is a setup recipe for running DeepSeek V4.1 Flash on two NVIDIA DGX Spark (GB10) machines. The latest release is v1.0.0. It runs the EXL3 2.9bpw model with TensorFold, and the README states that decoding is 2.1~2.3x faster than the vLLM setup. Prefill is the stage that reads in the input context. The README also covers tool calling and structured output, up to 4 concurrent requests, 1M-token context support, and how to connect an OpenAI API-compatible app.
The README's measurement table gives 27 tokens/s versus 57 tokens/s on prose tasks, and 40 versus 92 tokens/s on code tasks. Prefill of a 32,000-token prompt is 925 versus 2165 tokens/s. It states that the maximum context grew from 262,144 to 1,048,576 tokens. As evidence for these numbers, dated result files and run logs are published under runs/, 150 files in total. Output divergence tests, suitability tests, grammar checks, and benchmark runs are also recorded. The code is Python and the license is Apache-2.0.
The comparison target is not the standard vLLM but the vLLM kit, and the measurements are based on the README and the repository's run records. Because TensorFold is the author's own engine fork, independent verification cannot be confirmed from this material alone. The performance figures are therefore better read as an optimization case obtained on two DGX Spark machines, backed by reviewable measurement evidence, rather than as independently established results.