deepseek-v4.1THREAD
V4.1 Flash 1017 uses 18% fewer tokens on same pagoda prompt
Same prompt on two DGX Sparks: 1017 took 38,146 tokens and 16.9 min versus 46,355 tokens and 20.6 min for the older version, both at 37.6 tok/s.
Wësche
@WescheNex1qArchived 10 days agoOriginally posted 2026.10.18
THREAD
Same prompt, same 2 DGX Sparks, same speed. -DeepSeek V4.1 Flash: 46,355 tokens, 20.6 min -V4.1 Flash 1017: 38,146 tokens, 16.9 min 18% fewer tokens for the same pagoda. Both 37.6 tok/s, thinking on, no output cap. EXL3 2.9bpw quants by @Libertai_DAI (1017) @deepseek_ai
@kbsingh @techwithjoe_io @Libertai_DAI @deepseek_ai vLLM 179dd0fa9, but EXL3 isn't upstream. It's a community image that adds exllamav3 1.4.5 on top of vLLM, so building that hash alone won't load it https://ghcr.io/miaai-lab/deepseek-v4.1-flash-exl3-2x-dgx-sparks:2.9bpw
Wësche @WescheNex1qOriginally posted 2026.10.18
x.com/weschenex1q/status/2111589919806435647Archived 2026.10.18 16:27