deepseek-v4.1POST

Unsloth ships 1-bit GGUFs of DeepSeek V4.1-Flash-1017 at 233–250GB

The 1-bit quant is 233–250GB versus 508GB at Q8 and matches the full model's top token 85–88% of the time; the 233GB build runs on a 128GB Mac via an unmerged llama.cpp PR with its 111GB n-gram table on SSD.

kaimi

@kaimi_localArchived 11 days agoOriginally posted 2026.10.20

kaimi @kaimi_localOriginally posted 2026.10.20

Open original on X

Unsloth put out 1-bit GGUFs of DeepSeek V4.1-Flash-1017 at 233-250GB, vs 508GB at Q8, and they pick the full model's top token 85-88% of the time. The 233GB one runs on a 128GB Mac with an unmerged llama.cpp PR, which keeps its 111GB n-gram table on SSD.

This post quotes another post: x.com/unslothai/status/2112212605930778830

x.com/kaimi_local/status/2112367688169173181Archived 2026.10.20 12:02