Qwen3.5-9B Quantization New SOTA: 31 Wins, 0 Losses, KLD Improved 23%

KvAk_AKPlaysYT · reddit · 2026-08-14

Independent developer Aaryan Kapoor releases GGUF quantizations of Qwen3.5-9B, claiming 31 wins, 3 ties, 0 losses across 34 size-matched comparisons. His Q4KXL beats Unsloth's UD-Q4KXL by 23% in KLD while being smaller, and IQ2M is 36% closer to BF16 than UD-IQ2M at identical bytes, scoring 10 points higher on HumanEval+. All evals use the same BF16 reference. Author is an undergrad seeking internships in AI agent orchestration and inference.

Original post →

More from Models

Models channel →