4× DGX Spark 部署 V4.1 Flash 1017：2.9-bit 量化约 150 tok/s

Neko Legends @softpoo原推发布于 2026.10.18

在 X 上打开原推

DeepSeek dropped V4.1-Flash-1017 this morning. By afternoon it was serving on our 4× DGX Spark cluster — same TensorFold engine, no code changes, config-only drop-in. 2.9-bit: MMLU 89.5% (full-precision hosted: 90.0%). ~150 tok/s on code, 200K-token prompt read in 35.7s. Agentic 3D test (vision → build a garage in Blender over MCP): 108/180 — top of four models we've run: GLM-5.3-Flash 98, our previous DeepSeek 79, Nemotron 3.5 Super VL 44. Bonus: same weights, same 111 tasks — raw API, our pi harness, and DeepSeek's own harness all scored 101. The harness barely mattered.