REDDIT · REDDIT// 32d agoPRODUCT UPDATE

Unsloth refreshes Qwen3.5 GGUF lineup

Unsloth says its final Qwen3.5 GGUF refresh adds an improved quantization algorithm, new imatrix calibration data, and tool-calling fixes across key Qwen3.5 variants including 27B, 35B-A3B, 122B-A10B, and 397B-A17B. The update is positioned as a real quality pass for local inference, with refreshed benchmarks and a recommendation to re-download the affected models.

// ANALYSIS

This is the kind of update that matters more than a flashy model drop: better quants, fewer template bugs, and clearer performance tradeoffs for people actually running large models locally.

–Unsloth is optimizing for real workloads like chat, coding, long context, and tool calling, not just headline compression ratios
–The new imatrix data and quantization changes suggest the team is tuning for practical quality retention instead of blindly minimizing model size
–Retiring MXFP4 from several GGUF variants is notable because it admits some popular quant choices were hurting quality more than they helped
–Publishing detailed KL divergence benchmarks and large research artifacts makes this more credible than the usual opaque “improved weights” announcement

// TAGS

unslothqwen3.5llminferencebenchmarkopen-source

DISCOVERED

32d ago

2026-03-10

PUBLISHED

36d ago

2026-03-06

RELEVANCE

8/ 10

AUTHOR

jferments