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.
More from Models
- Open labs embrace continued post-training: GLM 5.3, Qwen3.8 27B show generational leaps without new base models — Daniel_H212 · 2026-08-14
- Qwen3.8-27B FP8 and Full Versions Available on Hugging Face — FrankWanders · 2026-08-14
- Grok 4.6 ranks #2 on EEBench, showing strong real-world electrical engineering skills — XFreeze · 2026-08-14
- Qwen3.8-27B tops Hugging Face trending, Apache 2.0 open weights — Qwen · 2026-08-14
- Qwen3.8-Max launches on Fireworks with Day-0 support: 2.4T-param MoE for agents and coding — Alibaba_Qwen · 2026-08-14
- Qwen3.8-27B Released: Checkpoint Loved by Startups and Single-Node Users — tokenbender · 2026-08-14