deepseek-v4.1IMAGE
DeepSeek V4.1 Flash EXL3 3.75bpw quant released for 3x DGX Spark
The new quant upgrades the highest-error gate/up layers to 4-bit, improving KL vs DeepSeek FP8 by 4–20% over 3.5bpw, and is on Hugging Face.
Thomas Eastham
@easthambuildsArchived 9 days agoOriginally posted 2026.10.22
Thomas Eastham @easthambuildsOriginally posted 2026.10.22
DeepSeek-V4.1-Flash-1017 EXL3 3.75 bpw for 3x DGX Spark, TensorFold TP3 This improves our 3.5 bpw quant plus routed expert bits: 4-bit down proj everywhere, 4-bit gate/up where 3-bit error was highest. KL vs DeepSeek's FP8: 4-20% better than 3.5 bpw https://huggingface.co/easthambuilds/DeepSeek-V4.1-Flash-1017-EXL3-3.75bpw
x.com/easthambuilds/status/2113278319554621708Archived 2026.10.23 12:08