deepseek-v4.1POST
单台 DGX Spark 跑 V4.1 Flash：35.5 tok/s
SAGE-EXL3 1.62bpw 量化 + vLLM 配方，单台 DGX Spark 解码 35.5 tok/s、预填充 1189 tok/s，HumanEval 与原生权重答案一致率 99.61%。
Neo
@NeoAIForecast11 天前收录原推发布于 2026.10.20
Neo @NeoAIForecast原推发布于 2026.10.20
One DGX Spark. DeepSeek V4.1 Flash 1017. 1.62 bits per expert. @ViC305 just dropped a SAGE-EXL3 quant and a one-Spark vLLM recipe!!! >35.5 tok/s decode >1,189 tok/s max prefill >229K context so far >99.61% answer-token agreement on HumanEval vs the native checkpoint weights = https://huggingface.co/vcruz305/DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw recipe = https://github.com/vcruz305/DSV4.1-Flash-1017-EXL3-Spark-recipe
x.com/neoaiforecast/status/2112386229564285133采集于 2026.10.20 12:02