---
license: mit
base_model:
- Qwen/Qwen3.5-27B
---
Optimized llama.cpp GGUF quants of Qwen 3.5 27b made for 16gb VRAM cards.

Made using [GGUF-Tool-suite!](https://github.com/Thireus/GGUF-Tool-Suite/) by Thireus, tuned by me :) (Gammaception)

Best 16 gb config for GC IQ3_M: (headless 64k ctx kv-cache@(q8_0,q8_0), -ub 256, q8_0 mmproj). Following official sampling parameters recommended, along with reasoning budget + message


EDIT 24/03/26: Updated GGUF to fix coding quality, new charts

EDIT 03/04/26: New variants with better quality ratios, update charts
| | unsloth ud-iq3_xxs | bartowski iq3_xs | GC-Thireus iq3_xs | mungert iq3_xs | bartowski iq3_m | GC-Thireus iq3_m | unsloth ud-q3_k_xl | bartowski iq4_xs |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Size (GiB) | 10.716 | 11.357 | 11.800 | 12.078 | 12.240 | 12.465 | 13.447 | 14.130 |
| PPL mean | 7.349445 | 7.275691 | 7.003592 | 7.146062 | 7.166074 | 6.886104 | 7.095870 | 6.932149 |
| KLD mean | 0.080486 | 0.066744 | 0.044320 | 0.053966 | 0.057891 | 0.032533 | 0.034783 | 0.019950 |
| Same top p | 88.546 | 89.837 | 91.560 | 90.625 | 90.782 | 92.926 | 92.656 | 94.610 |

kl logits from unsloth ud-q6_k_xl (best i can fit), wikitext-2 512ctx bf16 kv
![ppl_graph](charts5/pplMean.png "Chart showing PPL analysis of quants")
![kld_graph](charts5/kldMean.png "Chart showing KLD analysis of quants")
![tpp_graph](charts5/sameTopP.png "Chart showing Top P analysis of quants")