Qwen3.5 GGUF Benchmarks (unsloth.ai)

The Qwen3.5-35B model has undergone significant updates, with improvements in quantization techniques that exhibit state-of-the-art (SOTA) performance across various metrics. Recent benchmarks reveal that the model achieved a 99.9% KL Divergence score on the Pareto Frontier for Unsloth Dynamic quant formats, showcasing its effectiveness in compression without compromising performance. With over 9TB of research artifacts now available, the community can explore the detailed results of over 150 benchmarks, which include key insights into the impacts of different bit widths on model performance. A notable technical takeaway is the retirement of MXFP4 quantization methods in favor of alternatives like Q4_K, which demonstrates better performance on sensitive tensor types. While older methods faced significant degradation in certain configurations, the introduction of the Imatrix technique has shown promise in reducing KL Divergence by optimizing the quantization process. This reflects an ongoing shift in the AI community towards exploring more efficient quantization strategies that balance performance and resource usage, making these advancements critical for improving machine learning model deployments in real-world applications.