DeepSeek V4.1 Flash 763B

User reports about 34 tok/s decode with a 763B DeepSeek-V4.1-Flash-1017 build quantized for the Strata-based Xolotl engine. Setup is Xolotl (Strata) with a GSQ mixed-precision quant (Q3_K / Q2_0 / ternary per routed-expert projection) and FP4 KV cache at 512K context; the 153 GiB artifact is spread across VRAM, pinned RAM and NVMe, pulling in only the experts each token needs. The user says throughput is memory-bandwidth-bound and should land in the same class Xolotl advertises for V4.1; the build is clean, with no abliteration, and tuned for agentic and coding workloads.