---
title: Qwen3.8-27B-MFQ
canonical_url: "https://www.modelscope.cn/models/Tylogi/Qwen3.8-27B-MFQ"
md_url: "https://www.modelscope.cn/models/Tylogi/Qwen3.8-27B-MFQ.md"
repository: Tylogi/Qwen3.8-27B-MFQ
last_updated: 2026-08-24
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
base_model:
  - Qwen/Qwen3.8-27B
base_model_relation: quantized
downloads: 0
stars: 0
tags:
  - mfq
  - quantization
  - qwen
  - qwen3.8
---

# Qwen3.8-27B-MFQ

> Qwen3.8-27B-MFQ - Tylogi 在 ModelScope 开源的模型。Run Qwen3.8-27B with TyloQuant MFQ

Tylogi/Qwen3.8-27B-MFQ 是 ModelScope 魔搭社区上的text-generation模型，采用 apache-2.0 许可，基于 Qwen/Qwen3.8-27B 构建。

- **Repository**: Tylogi/Qwen3.8-27B-MFQ
- **License**: apache-2.0
- **Tasks**: text-generation
- **Base model**: Qwen/Qwen3.8-27B
- **Tags**: mfq, quantization, qwen, qwen3.8
- **Downloads**: 0
- **Stars**: 0
- **Last updated**: 2026-08-24

Source: https://www.modelscope.cn/models/Tylogi/Qwen3.8-27B-MFQ

---

## Run Qwen3.8-27B with TyloQuant MFQ

<p style="margin-top: 0; margin-bottom: 8px;">
  <em>TyloQuant MFQ provides neuron-anchored mixed-format quantization and direct CUDA/C++ inference for large models.</em>
</p>

<div style="display: flex; gap: 8px; align-items: center; flex-wrap: wrap;">
  <a href="https://github.com/Tylogi/TyloQuant">
    <img src="https://img.shields.io/badge/GitHub-TyloQuant-172033?logo=github&logoColor=white" alt="TyloQuant GitHub">
  </a>
  <a href="https://github.com/Tylogi/TyloQuant#readme">
    <img src="https://img.shields.io/badge/Documentation-English-2477d4" alt="English documentation">
  </a>
  <a href="https://github.com/Tylogi/TyloQuant/blob/master/README.zh-CN.md">
    <img src="https://img.shields.io/badge/Documentation-中文-2477d4" alt="中文文档">
  </a>
</div>

<div align="center">
  <a href="https://github.com/Tylogi/TyloQuant">
    <img src="https://github.com/Tylogi/TyloQuant/raw/master/docs/figures/tylogi-ai-lab.svg" width="520" alt="Tylogi AI Lab">
  </a>
</div>

<ul style="margin-top: 8px;">
  <li>This repository contains MFQ-quantized weights derived from the official <a href="https://modelscope.cn/models/Qwen/Qwen3.8-27B"><b>Qwen3.8-27B</b></a> release.</li>
  <li>Each <code>.mfq</code> file contains one complete model. Seven mixed VQ tiers and three scalar tiers are available.</li>
  <li>All listed results use the same complete WikiText-2 KLD evaluation against official BF16 reference logits.</li>
  <li>See the <a href="https://github.com/Tylogi/TyloQuant">project documentation</a> for runtime, format and deployment instructions.</li>
</ul>

### Available files and KLD

| MFQ file | Size | Mean KLD ↓ | Same-top ↑ |
| :--- | ---: | ---: | ---: |
| `Qwen3.8-27B-V1_S.mfq` | 5.765 GiB | 0.601539 | 67.7640% |
| `Qwen3.8-27B-V1_M.mfq` | 6.265 GiB | 0.388113 | 73.3584% |
| `Qwen3.8-27B-V2_XXS.mfq` | 6.757 GiB | 0.337361 | 75.0027% |
| `Qwen3.8-27B-V2_S.mfq` | 7.796 GiB | 0.176440 | 82.4037% |
| `Qwen3.8-27B-V3_XXS.mfq` | 10.183 GiB | 0.0611357 | 89.3097% |
| `Qwen3.8-27B-V3_S.mfq` | 11.214 GiB | 0.0428593 | 91.1927% |
| `Qwen3.8-27B-V4_XS.mfq` | 13.417 GiB | 0.0185307 | 94.0257% |
| `Qwen3.8-27B-S4_M.mfq` | 15.206 GiB | 0.00944699 | 95.7870% |
| `Qwen3.8-27B-S5_M.mfq` | 18.059 GiB | 0.00309604 | 97.4442% |
| `Qwen3.8-27B-S6_M.mfq` | 21.175 GiB | 0.00118449 | 98.4949% |

<p><sub>Evaluation protocol: WikiText-2 raw test split, no BOS token, context length 512, 580 chunks, 255 scored positions per chunk and 147,900 scored tokens. Reference logits come from the official Qwen3.8-27B BF16 weights. Mean KLD is lower-is-better; same-top is the fraction of scored positions whose top token matches the BF16 reference.</sub></p>

### Download

Install the current ModelScope client:

```bash
pip install -U modelscope
```

Download one tier without fetching the other model files:

```bash
modelscope download Tylogi/Qwen3.8-27B-MFQ Qwen3.8-27B-V2_S.mfq --local-dir ./Qwen3.8-27B-MFQ
```

Replace the filename with any tier listed above.

### License

The quantized weights follow the license of the original Qwen3.8-27B release: [Apache License 2.0](https://modelscope.cn/models/Qwen/Qwen3.8-27B/file/view/master/LICENSE).