The model was created from the original DeepSeek-V4-Flash-0731 GGUF and uses an experimental mixed quantization recipe designed to significantly reduce memory usage and increase inference speed on NVIDIA Blackwell GPUs.

Experimental quantization: This model has not been extensively evaluated for quality. Please benchmark it against the original model before using it for production workloads.

Base Model

DeepSeek-V4-Flash-0731

Original model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

License: MIT

This repository contains a derived/quantized GGUF representation of the original model.

Quantization Recipe

The model uses the following custom tensor-type recipe:

text

^token_embd\.weight$=Q8_0
^output\.weight$=Q8_0
^blk\.[0-9]+\.(attn|indexer).*\.weight$=Q6_K
^blk\.[0-9]+\.ffn_(gate|up|down)_shexp\.weight$=Q6_K
^.*hc_.*\.weight$=Q8_0
^blk\.(0|1|2|3|4|5|6|7|8|9|34|35|36|37|38|39|40|41|42)\.ffn_(gate|up|down)_exps\.weight$=IQ3_XXS
^blk\.(10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33)\.ffn_(gate|up|down)_exps\.weight$=IQ2_XXS

Oxidez/DeepSeek-V4-Flash-0731-UD-IQ3_IQ2-XS

DeepSeek-V4-Flash-0731 – Custom IQ3_XXS / IQ2_XXS GGUF

Base Model

Quantization Recipe

Oxidez/
DeepSeek-V4-Flash-0731-UD-IQ3_IQ2-XS