tl;dr: Unsloth released Dynamic v3.0 GGUF quantizations for Qwen3-VL-27B, claiming >10% better top-1 accuracy at the same file size versus other providers, using improved imatrix calibration, smarter layer selection, and pure post-training quantization (no QAT/QAD). New features include sub-1-bit quants (UD-IQ1_S at 6.2GB retaining ~72% top-1 accuracy) and a Divergence-300 @32 benchmark for evaluating multi-token trajectory fidelity against BF16. The quants work with llama.cpp and other standard inference engines.