Evaluation of Qwen3.8 27B Quantization Models: 4-Bit Performance and 1-Bit Degradation

The article evaluates Qwen3.8 27B quantizations, finding that 4-bit models maintain performance on benchmarks like GPQA Diamond and Terminal-Bench 2.1, while 1-bit models degrade significantly, performing near random chance. Cost and memory trade-offs are highlighted for different quantization levels.

Comments

Read original article