ResearchSep 9, 2026 · 19:53

Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses

From quesma.com

Image: quesma.com
AI;DR

Testing of Qwen3.8 27B quantizations using Unsloth GGUFs with llama.cpp found that the 4-bit Q4_K_M variant matches BF16 performance and fits an RTX 4090, while the 1-bit UD-IQ1_S variant collapsed.

Read the sourceSign in to suggest

Sources

+ 1 more
  1. DISCUSSIONmpweiherlobste.rs

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today