Kebob's picture
Update README.md
a0a02c7 verified
|
raw
history blame
939 Bytes
metadata
license: mit
base_model:
  - deepseek-ai/DeepSeek-V3-0324
tags:
  - ik_llama.cpp

ik_llama.cpp quantizations of DeepSeek-V3-0324

Quantized using ik_llama.cpp build = 3788 (4622fadc)

NOTE: These quants MUST be run using the llama.cpp fork, ik_llama.cpp

Credits to @ubergarm for his DeepSeek quant recipes for which these quants were based on.

name file size quant type bpw
DeepSeek-V3-0324-IQ4_KT 322.355 GiB IQ4_KT (97.5%) / Q8_0 (2.5%) 4.127
DeepSeek-V3-0324-IQ4_XS_R8 340.764 GiB IQ4_XS_R8 (97.5%) / Q8_0 (2.5%) 4.362
DeepSeek-V3-0324-D-IQ4_KS_R4 366.762 GiB IQ4_KS_R4 (65%) / IQ5_KS_R4 (32.5%) / Q8_0 (2.5%) 4.695
DeepSeek-V3-0324-D-Q4_K_R4 412.131 GiB Q4_K_R4 (65%) / Q6_K_R4 (32.5%) / Q8_0 (2.5%) 5.276
DeepSeek-V3-0324-Q8_0_R8 664.295 GiB Q8_0_R8 (100%) 8.504