[Daily Picks]

A LocalLLaMA write-up compared many Qwen3.8 27B GGUF/EXL3 quants on 16GB VRAM using Mean KLD against real C code—not just file size. The author's takeaway: bartowski IQ4_XS looked strong overall; ultra-aggressive tiny quants often underwhelmed; leave headroom for KV/context instead of maxing the weight file.

Practical tip: on 16GB, pick by quality-per-GB and usable context, then re-check with your own prompts. Chasing the smallest download is a common trap when the model still needs room to think.

Which Qwen3.8 27B quant are you running on 16GB, and did smaller always mean worse for you?

Source: https://www.reddit.com/r/LocalLLaMA/comments/1w7ee1c/i_benchmarked_21_qwen38_27b_variants_on_16gb_vram/