New todaydeepseek-v4.1POST

DeepSeek V4.1 Flash on RTX PRO 6000: ~77 tok/s on code at 16K context

Running DeepSeek V4.1 Flash locally on an RTX PRO 6000 96GB with mixed IQ1/IQ2 experts and speculative decoding gives about 70 tok/s on prose and 77 on code at 16K context.

ideahazard💹🧲

@1d34h4z4rdArchived 8 hours agoOriginally posted 2026.10.28

ideahazard💹🧲 @1d34h4z4rdOriginally posted 2026.10.28

Open original on X

@0xSero DeepSeek V4.1 Flash locally: RTX PRO 6000 Blackwell 96GB + 128GB RAM, Hermes/Strata. Mixed IQ1/IQ2 experts, IQ4 Engram + DSpark speculative decoding. At 16K context: ~70 tokens/sec prose, ~77 code, 218–224 on easy sequences. Stock + abliterated. Works well but ill bench Qwen too

x.com/1d34h4z4rd/status/2115213390482145294Archived 2026.10.28 12:02