@sudoingX’s Two-Spark DeepSeek Recipe v2.0 Claims the Closest Quant Yet

A vLLM route for two Sparks, with every number in the repo.

@sudoingX’s recipe runs DeepSeek-V4.1-Flash-1017 on two DGX Sparks with his own EXL3 weights, made from the official checkpoint, on upstream vLLM with his patches and decode kernels. Release v2.0 came out Oct 22, he says, with a 204 GB build averaging 2.80 bpw. The context window is 262k; vision is in but off by default, since it costs about a gigabyte per Spark.

The official checkpoint is 510 GB, his README notes: 289 GB of MXFP4 routed experts, 203 GB of Engram tables and 18.5 GB of everything else. The recipe re-encodes the experts to about 191 GB and reads the Engram tables from NVMe at lookup time.

Of the five 1017 packs he measured, his README says v2.0 is closest to native on both test panels. HeNryous’s 2.75 bpw build is next at 0.0201. TensorFold keeps more free memory and serves a 1M window and four sequences, against his 262,144 tokens and two.

An outside run, under different settings.

@niklaslenz_ai ran the same v2.0 release, commit f8f20b3, with DeepSeek’s card sampling of temperature 1.0. He measured 64.9 tok/s on code, 62.2 on prose and 66.3 tok/s in total with eight requests. It scored 71.3% on MMLU-Pro and 90.0% on IFEval, and built a voxel world in OpenCode in 49.3 minutes. The two code figures come from different sampling, so they are not a like-for-like comparison.