Benchmarking Qwen3.8 27B Across 9 KV Quantization Methods and 189 Tests

On October 15, 2026, a user posted on the LocalLLaMA subreddit under the title "Tested Qwen3.8 27B: 189 benchmark results across 9 KV quantizations." According to the post, the author calculated KLD and perplexity for Qwen3.8 27B quantization combinations that fit within 16GB of VRAM, covering 9 KV quantization types, 21 models, and 189 benchmarks in total. The post notes that the benchmark data was placed in the comments because Reddit's filters kept deleting the submission.

The stated conclusion is that q5_1/q5_1 is almost always better than q8_0/q4_0 or q4_0/q8_0. The difference between q8_0/q4_0 and q4_0/q8_0 is described as very small, with q4_0/q8_0 sometimes outperforming q8_0/q4_0, so the author suggests it is safer to simply go with q5_1/q5_1. The post also says that moving up one model quantization and running at q4_0/q4_0 is almost always better in terms of KLD than running the quantization below at f16/f16.

So far the only available account is this single Reddit post. It was published by an individual user, concerns Qwen3.8 27B, a 16GB VRAM constraint, 9 KV quantization types, 21 models, and 189 benchmarks, and uses KLD and perplexity as its metrics. The post body does not include the full results and instead points readers to the comments.