deepseek-v4.1IMAGE

双 DGX Spark 部署 V4.1 Flash：代码 111.5 tok/s

因单台 DGX Spark 装不下，用 TensorFold TP=2 拆到两台，单流代码 111.5 tok/s，比 Qwen3.8-Flash-Next 快约 20%。

Thor

@coolthor15 小时前收录原推发布于 2026.10.28

双 DGX Spark 部署 V4.1 Flash：代码 111.5 tok/s

Thor @coolthor原推发布于 2026.10.28

DeepSeek V4.1 Flash doesn't fit on one DGX Spark, so I split it across two with TensorFold TP=2. Single-stream it's ~20% faster than Qwen3.8-Flash-Next (111.5 vs 92.4 tok/s on code). 8-stream Chinese is on par. Setup and gotchas: https://ai-muninn.com/en/blog/deepseek-v41-flash-dual-dgx-spark-tensorfold

x.com/coolthor/status/2115141559771390352采集于 2026.10.28 12:03