---

base_model:

• Qwen/Qwen3.8-27B

license: apache-2.0

tags:

• unsloth
• imatrix
• llama.cpp
• qwen3.8
• qwen
• ubergarm
• 16GB

---

Qwen3.8-27B (GGUF — ZB-ZipBrain Quantization)

This repository provides GGUF quantizations for Qwen3.8-27B optimized using ZB-ZipBrain, a layer-wise quantization profiling and allocation method.

1. Overview & Method: ZB-ZipBrain

ZB-ZipBrain is an automated layer-allocation approach that dynamically profiles model layers and mixes K-quants and IQ-quants based on layer sensitivity and importance matrices.

> I developed and tested this method alongside AI over the past three days. It was created purely for research purposes 😊

Key Objectives:

• Selective Bit-rate Allocation: Assigns higher precision to sensitive layers and compact IQ-quants to more resilient weights.
• Balanced Efficiency: Maintains low perplexity (PPL) and minimal Kullback-Leibler (KL) Divergence relative to the BF16 baseline while achieving target file sizes / bits-per-weight (bpw).

2. Benchmark & Evaluation Results

All models were evaluated against the BF16 baseline (Mean PPL = 6.950493) using standard Perplexity (PPL) and KL Divergence metrics.

Comprehensive Comparison Table

Updated ranked list (re-sorted primarily by Mean KLD ascending — lower is better; ties broken by other quality metrics):

| Rank | Model / File Name | Level / Source | Quant Type | Size (GB) | Mean PPL | Δ PPL | Mean KLD | Same Top-p (%) | KLD 99% |

| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |

| 1 | Qwen3.8-27B-UD-Q8_K_XL | unsloth UD2 | UD-Q8_K_XL | 29.30 | 6.953800 | +0.003500 | 0.000850 | 98.970% | — |

| 2 | Qwen3.8-27B-UD-Q6_K_XL | unsloth UD2 | UD-Q6_K_XL | 24.14 | 6.953600 | +0.003200 | 0.001380 | 98.520% | — |

| 3 | Qwen3.8-27B-Q6_K | unsloth UD2 | Q6_K | 21.31 | 6.950700 | +0.000300 | 0.002290 | 97.860% | — |

| 4 | Qwen3.8-27B-Q5_K_M | unsloth UD2 | Q5_K_M | 18.47 | 6.974200 | +0.023900 | 0.006220 | 96.700% | — |

| 5 | Qwen3.8-27B-UD-Q4_K_XL | unsloth UD2 | Q4_K_XL | 16.69 | 6.979220 | +0.028728 | 0.008606 | 96.091% | 0.091099 |

| 6 | Qwen3.8-27B-ZB4.97-GOD-IQ4_XS | ZB-GOD | IQ4_XS | 15.82 | 7.004243 | +0.053751 | 0.012249 | 95.337% | 0.117617 |

| 7 | Qwen3.8-27B-UD3-Q4_K_S | unsloth UD3 | Q4_K_S | 14.30 | 6.969514 | +0.019022 | 0.013652 | 95.149% | 0.141744 |

| 8 | Qwen3.8-27B-Autoround-Q4_K_M | intel | Q4_K_M | 15.66 | 6.950294 | -0.000199 | 0.014657 | 94.859% | 0.147949 |

| 9 | Qwen3.8-27B-ZB4.65-PRO-IQ4_XS | ZB-PRO | IQ4_XS | 14.81 | 7.017278 | +0.066786 | 0.015466 | 94.766% | 0.150252 |

| 10 | Qwen3.8-27B-Q4_K_M | unsloth UD2 | Q4_K_M | 15.93 | 6.956100 | +0.005800 | 0.015490 | 94.650% | — |

| 11 | Qwen3.8-27B-ZB4.60-PRO-IQ4_XS | ZB-PRO | IQ4_XS | 14.65 | 7.030895 | +0.080402 | 0.016162 | 94.668% | 0.159115 |

| 12 | Qwen3.8-27B-ZB4.55-PRO-IQ4_XS | ZB-PRO | IQ4_XS | 14.49 | 7.032263 | +0.081771 | 0.016647 | 94.613% | 0.161016 |

| 13 | Qwen3.8-27B-IQ4_NL | bartowski | IQ4_NL | 15.20 | 7.006472 | +0.055980 | 0.018427 | 94.230% | 0.190168 |

| 14 | Qwen3.8-27B-IQ4_XS | unsloth UD2 | IQ4_XS | 14.63 | 7.012695 | +0.062202 | 0.018652 | 94.270% | 0.194338 |

| 15 | Qwen3.8-27B-UD3-IQ4_XS | unsloth UD3 | IQ4_XS | 13.27 | 7.004732 | +0.054240 | 0.018772 | 93.975% | 0.195164 |

| 16 | Qwen3.8-27B-ZB4.48-STD-IQ4_XS | ZB-STD | IQ4_XS | 14.26 | 7.050096 | +0.099604 | 0.018892 | 94.199% | 0.196005 |

| 17 | Qwen3.8-27B-Q4_K_S | unsloth UD2 | Q4_K_S | 15.01 | 6.966826 | +0.016334 | 0.018921 | 94.235% | 0.192749 |

| 18 | Qwen3.8-27B-IQ4_XS-i1 | mradermacher | IQ4_XS | 14.26 | 7.012810 | +0.062318 | 0.019271 | 94.141% | 0.197891 |

| 19 | Qwen3.8-27B-Q4_K_S-i1 | mradermacher | Q4_K_S | 14.74 | 6.989551 | +0.039059 | 0.019805 | 93.996% | 0.204143 |

| 20 | Qwen3.8-27B-ZB4.36-STD-IQ4_XS | ZB-STD | IQ4_XS | 13.88 | 7.054811 | +0.104319 | 0.020556 | 93.951% | 0.206689 |

| 21 | Qwen3.8-27B-Q4_0-AutoRound-Code | webhie | Q4_0 | 14.64 | 7.067142 | +0.116650 | 0.026586 | 92.970% | 0.271619 |

| 22 | Qwen3.8-27B-ZB4.14-MIN-IQ4_XS | ZB-MIN | IQ4_XS | 13.19 | 7.045689 | +0.095196 | 0.029334 | 92.799% | 0.294996 |

| 23 | Qwen3.8-27B-IQ4_XS-Smaller_3.96 | jrell | IQ4_XS | 12.61 | 7.252766 | +0.302274 | 0.055499 | 90.090% | 0.551972 |

Notes on the new entries :

• UD3-Q4_K_S (rank 7): Excellent KLD and Same Top-p for its size; strong contender among ~14 GB models.
• Autoround-Q4_K_M (rank 8): Very close to base PPL (slightly better ΔPPL), solid KLD.
• UD3-IQ4_XS (rank 15): Competitive with other IQ4_XS variants, good size/quality trade-off.

Update: Aug 20, 2026

>The newly released Unsloth Dynamic v3 is truly the best value for performance right now.

>My ZB is just an experiment, feel free to check it out for fun :)

3. ZB-ZipBrain Tiers & Recommendations

• ZB-GOD : God. A singularity appears. Reaches 0.012249 Mean KLD and 95.34% top-probability match.
• ZB-PRO : Pro. Recommended for 16 GB VRAM GPUs with offloading on CPU. Balances quality output with substantial size savings.
• ZB-STD : Standard. Similar to other standard IQ4_XS models currently available.
• ZB-MIN : Minimal. Optimal footprint for tight 16 GB memory setups, allowing headroom for longer context windows

---

Credits & Acknowledgements

• Base Model: Qwen3.8 27B by Alibaba Cloud / Qwen Team.
• BF16 Base GGUF: Provided by Unsloth AI.
• Importance Matrix (imatrix): Generated and curated by ubergarm.
• Inference & Quantization Framework: llama.cpp by Georgi Gerganov and contributors.