DeepSeek V4.1 Flash 1017 on vLLM + EXL3 runs at only 1 tok/s

A user serving DeepSeek V4.1 Flash 1017 with ViC305 quantization and vllm-exl3 got it loading stably after a few hours, but throughput is just 1 tok/s; root cause still being investigated.

Hassan Emam

@hassanemamArchived 9 days agoOriginally posted 2026.10.23

Hassan Emam @hassanemamOriginally posted 2026.10.23

Open original on X

I am experimenting serving deepseek V4.1 Flash 1017 using @ViC305 quantization and vllm-exl3. After few hours I managed to load the model and make it stable. The bad news it is running on 1 tok/s. I am still working on finding the root cause. Stay tunned

x.com/hassanemam/status/2113325247533363247Archived 2026.10.23 12:08

Open original on X