Tested @UnslothAI new Dynamic V3 GGUFs on a 4090. They deliver a Q3 that beats a Q4 4.5GB bigger. All runs: same box, same harness, wikitext-2 (60 chunks), llama.cpp + DFlash2 speculative decoding, q4_0 KV, 262k ctx. Quant | Size | PPL | Med t/s | Needle retrieval 🏆Unsloth UD-Q3 12.24GB 6.3993 110.7 250k tok IQ4_XS 15.1GB 6.4149 107.3 32k Unsloth UD-Q4 16.7GB 6.4181 62.6 4k My Q3 (imatrix) 12.57GB 6.5316 84.8 258k tok My hand-tuned 13.1GB 6.5865 94.0 120k Takeaways: • Their Q3 has the best PPL of anything I tested including Q4s • Smallest file, fastest median, and it retrieved an exact needle at 250k tokens • I tried hand-rolling my own layer mix (q6_K embeddings, q5_K attention). It was the worst result. The recipe is the value per-layer types derived from error analysis • Their Q2 does degrade (6.6469, +3.8%) — Q3 is the floor where quality holds Dynamic V3: