license: apache-2.0
base_model: Qwen/Qwen3.8-27B
pipeline_tag: image-text-to-text
library_name: mlx
tags:
- qwen
- qwen3.8
- mlx
- mlx-vlm
- safetensors
- abliterix
- quantized
- multimodal
- vision-language
language: - multilingual
Qwen3.8-27B Abliterated — MLX Collection
[!IMPORTANT]
This is an unofficial community derivative of Qwen/Qwen3.8-27B. It has been processed with Abliterix to reduce selected refusal-related behavior and converted to MLX-compatible Safetensors in 8-bit, 6-bit, and 4-bit variants.This is not an official Qwen release and is not endorsed by the Qwen Team, Alibaba Cloud, Apple, the MLX/MLX-VLM maintainers, or the Abliterix authors.
本仓库是 Qwen/Qwen3.8-27B 的非官方社区衍生版本。模型经过 Abliterix 处理，以降低部分拒答相关行为，并转换为适用于 MLX 生态的 8bit、6bit 和 4bit Safetensors 权重。
本模型并非 Qwen 官方发布，也不代表 Qwen 团队、阿里云、Apple、MLX/MLX-VLM 维护者或 Abliterix 作者的立场或认可。
Model Summary / 模型简介
Qwen3.8-27B is a dense, approximately 27-billion-parameter native vision-language model supporting text, image, and video understanding. This repository preserves the derivative model architecture while applying a refusal-direction intervention with Abliterix and MLX affine weight quantization.
Qwen3.8-27B 是一个约 270 亿参数的原生视觉语言模型，支持文本、图像和视频理解。本仓库在保留衍生模型架构的基础上，通过 Abliterix 对拒答相关方向进行干预，并使用 MLX affine 权重量化提供不同精度版本。
Repository relationship / 仓库关系
- Official base model / 官方基础模型：
Qwen/Qwen3.8-27B - Behavioral intervention / 行为干预： Abliterix refusal-direction intervention
- Distribution format / 分发格式： MLX-compatible sharded Safetensors
- Quantization mode / 量化模式： affine, group size 64
- Available variants / 可用版本： 8bit, 6bit, 4bit
- Repository / 仓库：
GlobalCybersecurityAlliance/Qwen3.8-27B-Abliterated-MLX - Maintainer / 维护者： Global Cybersecurity Alliance (community distribution)
The official Qwen benchmark results describe the unmodified base model only. They must not be interpreted as benchmark results for this Abliterix-processed or MLX-quantized derivative. No claim is made that this derivative preserves every capability or score of the official model.
官方 Qwen 模型卡中的基准测试结果仅代表未经修改的基础模型，不能视为本 Abliterix 衍生版或 MLX 量化版的实测成绩。本仓库不保证衍生模型完整保持官方模型的全部能力或分数。
Available Variants / 可用版本
| Folder | Quantization | Group size | Approx. local size | Suggested use |
|---|---|---|---|---|
8bit/ |
8-bit affine | 64 | 28GB | Highest fidelity in this collection; highest memory requirement |
6bit/ |
6-bit affine | 64 | 22GB | Recommended balance between quality, memory use, and storage |
4bit/ |
4-bit affine | 64 | 15GB | Lowest memory requirement; quality may differ more from the BF16 derivative |
The sizes above are approximate filesystem sizes and may differ slightly from the values shown by Hugging Face. Each folder is a self-contained model directory and includes the model configuration, tokenizer, processor configuration, chat template, sharded weights, and weight index.
以上大小是本地文件系统中的近似值，可能与 Hugging Face 页面显示值略有差异。每个目录都是独立、完整的模型目录，包含模型配置、分词器、处理器配置、聊天模板、分片权重及权重索引。
QUANTIZATION_COMPLETE.json, when present, is build-verification metadata and is not required for inference.
What Abliterix Changes / Abliterix 修改说明
Abliterix identifies and modifies activation directions associated with refusal behavior. The intended effect is to reduce excessive refusals on selected evaluations. This is a weight-level behavioral intervention rather than a system-prompt override.
Abliterix 用于识别并修改与拒答行为相关的激活方向，目标是在特定评测中减少过度拒答。这属于权重层面的行为干预，并非简单覆盖系统提示词。
Important limitations / 重要限制：
- Reduced refusal behavior is not guaranteed for every prompt, language, chat template, or inference engine.
- The model may still refuse requests because refusal behavior can be distributed across multiple layers and mechanisms.
- Abliterix processing may affect tone, calibration, reasoning quality, factual accuracy, safety behavior, or instruction following.
- Quantization may introduce additional quality differences compared with the BF16 derivative.
- The internal label
best15is a build-selection label, not a standardized or independently reproduced benchmark score. - “Abliterated” or reduced refusal does not mean unrestricted capability, guaranteed compliance, correctness, or safety.
- 模型不保证对所有提示词、语言、聊天模板或推理框架都降低拒答。
- 拒答行为可能分布于多个层和机制，因此模型仍可能拒绝部分请求。
- Abliterix 处理可能影响语气、置信度校准、推理质量、事实准确性、安全行为或指令遵循能力。
- 量化可能进一步造成与 BF16 衍生模型不同的质量变化。
- 内部标签
best15仅用于构建版本筛选，不是标准化或独立复现的评测分数。 - “Abliterated”或拒答率降低不代表模型具备无限能力，也不保证服从、正确或安全。
Requirements / 运行要求
Use a recent version of MLX-VLM with support for the architecture declared by the included config.json. Compatibility depends on the installed MLX-VLM/Transformers version and the available backend.
建议使用最新版 MLX-VLM，并确保其支持本模型 config.json 声明的架构。实际兼容性取决于 MLX-VLM、Transformers 版本及可用计算后端。
python -m pip install -U mlx-vlm huggingface_hub
These models are large. Available memory must exceed the weight size and should leave additional headroom for the vision encoder, runtime buffers, prompt processing, and KV cache. Longer contexts and larger images require more memory. Start with the 4bit version and a short context if resources are limited.
本模型体积较大。可用内存不仅需要容纳模型权重，还应为视觉编码器、运行时缓冲区、提示词处理和 KV Cache 预留空间。上下文越长、图像越大，内存需求越高。资源有限时建议先使用 4bit 版本并缩短上下文。
Download a Variant / 下载指定版本
Because all three variants are stored as subfolders in one repository, download the required folder first. The examples below use the recommended 6bit version.
由于三个版本存放在同一仓库的不同子目录中，请先下载所需目录。以下示例使用推荐的 6bit 版本。
hf download GlobalCybersecurityAlliance/Qwen3.8-27B-Abliterated-MLX \
--include "6bit/*" \
--local-dir ./Qwen3.8-27B-Abliterated-MLX
The local model path will then be:
./Qwen3.8-27B-Abliterated-MLX/6bit
Replace 6bit with 8bit or 4bit in both places to download another variant.
Quick Start with MLX-VLM / 使用 MLX-VLM
The following commands follow the current MLX-VLM CLI style. Argument names and architecture support may change across releases; run mlx_vlm.generate --help or mlx_vlm.server --help when necessary.
以下命令采用当前 MLX-VLM 的命令行格式。不同版本的参数名称和架构支持可能变化，如遇问题请运行 mlx_vlm.generate --help 或 mlx_vlm.server --help。
Text generation / 文本生成
mlx_vlm.generate \
--model ./Qwen3.8-27B-Abliterated-MLX/6bit \
--prompt "Explain the difference between symmetric and asymmetric encryption." \
--max-tokens 512 \
--temperature 0.6
Image understanding / 图像理解
mlx_vlm.generate \
--model ./Qwen3.8-27B-Abliterated-MLX/6bit \
--image /path/to/image.jpg \
--prompt "Describe this image in detail." \
--max-tokens 512 \
--temperature 0.2
Start an API server / 启动 API 服务
mlx_vlm.server \
--model ./Qwen3.8-27B-Abliterated-MLX/6bit \
--port 8080
Check the server:
curl http://localhost:8080/health
The MLX-VLM server is suitable for experimentation. Before exposing any endpoint to a network, configure authentication, network isolation, access control, logging, rate limits, and other safeguards appropriate to the deployment.
MLX-VLM 服务适合实验用途。在将接口暴露到网络之前，请根据实际环境配置身份认证、网络隔离、访问控制、日志审计、速率限制及其他必要保护措施。
Choosing a Variant / 如何选择版本
- Choose 8bit when output quality is the priority and sufficient memory is available.
- Choose 6bit as the recommended general-purpose balance.
- Choose 4bit when memory and storage are constrained.
- 质量优先且内存充足时选择 8bit。
- 一般用途推荐选择质量和资源占用较均衡的 6bit。
- 内存或存储空间有限时选择 4bit。
Quantization level does not directly determine refusal rate. Sampling parameters such as temperature and top-p can change output variability, but they do not reproduce or replace the weight-level Abliterix intervention.
量化位宽并不直接决定拒答率。temperature、top-p 等采样参数会改变输出随机性，但不能复现或替代 Abliterix 的权重层干预。
Intended Use / 预期用途
Appropriate uses include:
- research into model behavior, alignment, refusal, robustness, and quantization;
- authorized evaluation and red-team testing;
- local MLX/MLX-VLM deployment experiments;
- defensive cybersecurity education and research in authorized environments;
- evaluation of multimodal model behavior on text, image, and video inputs.
适合的用途包括：模型行为、对齐、拒答、鲁棒性与量化研究；经授权的评估和红队测试；MLX/MLX-VLM 本地部署实验；合法授权环境中的防御性网络安全教育与研究；文本、图像和视频多模态行为评估。
Prohibited and High-Risk Use / 禁止及高风险用途
Do not use this model to facilitate unlawful activity, unauthorized access, credential theft, malware deployment, privacy invasion, harassment, violence, fraud, exploitation, or other harm. Operators must implement access controls, monitoring, rate limits, content safeguards, and human review appropriate to the deployment context.
不得使用本模型实施或协助违法活动、未经授权的系统访问、凭据窃取、恶意软件投放、侵犯隐私、骚扰、暴力、欺诈、剥削或其他伤害行为。部署者必须根据应用场景实施访问控制、审计监控、速率限制、内容保护和人工复核。
Disclaimer / 免责声明
[!CAUTION]
USE AT YOUR OWN RISK. Abliterix processing intentionally changes refusal-related behavior and may weaken safeguards present in the official model. The model may generate inaccurate, unsafe, offensive, biased, unlawful, insecure, or otherwise harmful content and may follow malicious instructions more readily than the official base model.
This repository and all included files are provided “AS IS” and “AS AVAILABLE,” without warranties or conditions of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, accuracy, reliability, non-infringement, security, safety, compatibility, performance, or uninterrupted availability.
The maintainers, contributors, quantizers, converters, distributors, the Qwen Team, Alibaba Cloud, Apple, MLX/MLX-VLM maintainers, and the Abliterix authors are not responsible for prompts, outputs, decisions, deployments, damages, losses, security incidents, claims, liabilities, or consequences arising from use or misuse of this derivative, to the maximum extent permitted by applicable law. Mention of any organization or software project does not imply endorsement or involvement.
Users and deployers are solely responsible for:
- evaluating the model before deployment;
- complying with applicable laws, regulations, licenses, platform policies, and third-party rights;
- obtaining explicit authorization before any cybersecurity testing;
- preventing access by unauthorized or unsuitable users;
- implementing safeguards appropriate to the use case;
- independently verifying model outputs before relying on them;
- protecting personal, confidential, proprietary, and security-sensitive data.
The model must not be treated as professional medical, legal, financial, safety, security, cybersecurity, or operational advice. Do not rely on it for decisions where errors could cause injury, rights violations, financial loss, system compromise, or other material harm.
[!CAUTION]
使用者自行承担全部风险。 Abliterix 处理会主动改变拒答相关行为，可能削弱官方模型原有的部分安全保护。模型可能生成错误、不安全、冒犯性、偏见性、违法、不可靠或其他有害内容，也可能比官方基础模型更容易遵循恶意指令。
本仓库及其中所有文件均按**“现状”和“可用状态”**提供，不作任何明示或默示保证，包括但不限于适销性、特定用途适用性、准确性、可靠性、不侵权性、安全性、无害性、兼容性、性能或持续可用性保证。
在适用法律允许的最大范围内，维护者、贡献者、量化者、转换者、分发者、Qwen 团队、阿里云、Apple、MLX/MLX-VLM 维护者以及 Abliterix 作者，均不对因使用或误用本衍生模型产生的提示词、输出、决策、部署、损害、损失、安全事件、索赔、责任或后果承担责任。文中提及任何组织或软件项目，均不代表其认可或参与本模型。
使用者和部署者应自行负责：
- 在部署前充分评估模型；
- 遵守适用法律法规、许可证、平台规则和第三方权利；
- 在开展任何网络安全测试前取得明确授权；
- 防止未经授权或不适合的用户访问；
- 根据实际用途实施必要的安全保护措施；
- 在依赖任何模型输出前进行独立核验；
- 保护个人、机密、专有及安全敏感数据。
本模型不得被视为专业的医疗、法律、金融、安全、网络安全或运营建议。对于错误可能导致人身伤害、权利侵害、经济损失、系统失陷或其他重大损害的决策，不得直接依赖模型输出。
License and Attribution / 许可证与署名
The official base-model repository identifies its license as Apache-2.0. Users must independently review and comply with the base-model license and any terms applicable to dependencies, MLX/MLX-VLM, inference software, datasets, inputs, outputs, and the intended use.
This derivative does not transfer ownership of the original model, trademarks, documentation, software, or third-party materials. “Qwen,” related marks, and official documentation remain the property of their respective owners.
官方基础模型仓库将许可证标注为 Apache-2.0。使用者必须自行查阅并遵守基础模型、依赖项、MLX/MLX-VLM、推理软件、数据集、输入、输出及实际用途适用的许可证和相关条款。
本衍生版本不转移原始模型、商标、文档、软件或任何第三方材料的所有权。“Qwen”及相关标识和官方资料仍归各自权利人所有。