Atomic releases 14 DeepSeek-V4-Flash GGUF quants, recommends AD-IQ2_M for 128GB rigs

testingcatalog · x · 2026-08-04

Atomic released 14 GGUF quantizations of DeepSeek-V4-Flash 0731 on Hugging Face, spanning lossless BF16 down to 1-bit.

The model is a 284B-parameter MoE checkpoint trained with quantization awareness; the image compares 38 GGUF variants from 7 publishers and highlights AD-IQ2M as the best fit for 128GB hardware, with token-choice agreement reported at 83.6% versus the community’s other V4 Flash GGUFs.

Related event: Atomic Releases 14 Quantized Versions of DeepSeek V4 Flash(2 posts)→

Original post →

More from Infra

Infra channel →