← back to catalog · registered 2026-10-20 00:00

NeosQ/Qwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed

Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/NeosQ%2FQwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed"
Response includes
  • classification m-uncensored
  • files 37
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M-U
Primary method

Uncensored (method unknown)

No other method signals detected in this model.
Confidence
LOW
Why this label 3 signals
Weak or ambiguous signals. Best guess based on catalog patterns; treat as tentative and check the evidence below.
  • 'uncensored' in name/tags but no 'abliterated' marker
  • method not identifiable from author declaration alone
  • may be DPO fine-tune, prompt engineering, or unknown technique
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-10-19

Metadata

License
apache-2.0
Languages
en zh
Tags
mlx safetensors qwen4_exp mtplx mtp speculative-decoding apple-silicon uncensored abliterated en zh license:apache-2.0
Total size
107 GB
Files
37
Quantizations
1
Registered
2026-10-20 00:00
Last updated on HF
2026-10-19 23:54

Files by quantization

Auxiliary files 37 files 107 GB
ngram-table.safetensors 29.8 GB db2b050c download
model-00012-of-00020.safetensors 4.00 GB 1ce99628 download
model-00006-of-00020.safetensors 4.00 GB 493ada62 download
model-00014-of-00020.safetensors 4.00 GB 5ec1924f download
model-00004-of-00020.safetensors 4.00 GB d0b6202a download
model-00010-of-00020.safetensors 3.96 GB b766ccd4 download
model-00016-of-00020.safetensors 3.96 GB 6435a8f8 download
model-00018-of-00020.safetensors 3.96 GB 553856d7 download
model-00008-of-00020.safetensors 3.96 GB ae14874d download
model-00002-of-00020.safetensors 3.96 GB fa075172 download
model-00001-of-00020.safetensors 3.80 GB 67693bae download
model-00011-of-00020.safetensors 3.74 GB 209346d1 download
model-00019-of-00020.safetensors 3.74 GB ce4ae8ef download
model-00003-of-00020.safetensors 3.74 GB af7a5904 download
model-00017-of-00020.safetensors 3.73 GB 911a92ff download
model-00009-of-00020.safetensors 3.73 GB 94ed4148 download
model-00005-of-00020.safetensors 3.69 GB 7256020a download
model-00015-of-00020.safetensors 3.69 GB 5ba89be0 download
model-00007-of-00020.safetensors 3.69 GB 2975158b download
model-00013-of-00020.safetensors 3.69 GB ab191847 download
model-00020-of-00020.safetensors 2.14 GB fa748a46 download
mtp.safetensors 1.58 GB f5beffff download
model-vision.safetensors 856 MB 5aff6983 download
tokenizer.json 20.0 MB 4de83bcb download
vocab.json 6.41 MB 0aa0ce06 download
merges.txt 3.20 MB a494e019 download
model.safetensors.index.json 278 KB 733e7c25 download
config.json 79.1 KB 994db35b download
tokenizer_config.json 17.5 KB 5de744b3 download
chat_template.jinja 8.74 KB c0c686f9 download
README.md 5.63 KB 177bc614 download
mtplx_runtime.json 4.70 KB 4181ff71 download
.gitattributes 1.53 KB 52373fe2 download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
processor_config.json 367 B ced889f7 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


license: apache-2.0
language:

  • en
  • zh
    tags:
  • mtplx
  • mlx
  • mtp
  • speculative-decoding
  • apple-silicon
  • uncensored
  • abliterated

Qwen3.8-Flash-Next-Uncensored · MTPLX Optimized Speed

English summary: This is an Apple Silicon (MLX/MTPLX) quantization of
orcarouter/Qwen3.8-Flash-Next-Uncensored —
OrcaRouter's abliterated (refusal-removed) BF16 build of Qwen's Qwen3.8-Flash-Next (125B-A6B MoE with a native MTP
speculative-decoding head). Quantization was performed locally with MTPLX Forge using the official
flash-next-optimized-speed recipe (4-bit body + 8-bit sparse-attention projections), the same recipe as
Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed.
The native MTP head is preserved and verified: 2.07x faster than plain decoding in Forge verification on an M5 Max 128 GB.
Requires MTPLX on Apple Silicon (96–128 GB+ unified memory).


这是什么

这是 OrcaRouter 去审查权重的 MTPLX 量化版。具体来说:

  • 权重来源:orcarouter/Qwen3.8-Flash-Next-Uncensored
    —— OrcaRouter 对 Qwen/Qwen3.8-Flash-Next(125B 参数 / 约 6B 激活的 MoE,混合注意力,原生视觉,自带 MTP
    投机解码头)做的消融式去审查(abliteration)BF16 版本
  • 量化制作:本仓库的量化并非 Orca 官方发布,而是使用 MTPLX Forge(v2.12.2)在本地从上述 BF16
    权重构建,采用 MTPLX 官方内置配方 flash-next-optimized-speed —— 与
    Youssofal 官方包完全相同的配方
  • 运行环境:MTPLX(Apple Silicon 专属,macOS 14+),
    利用模型自带的 MTP 头做精确投机采样加速

量化配置(q4 + q8 注意力)

组件 精度 说明
模型主体(专家、投影等) 4-bit affine, group 32 动态混合精度
稀疏注意力投影(QSA) 8-bit, group 64 保质量的关键部位
MTP 投机解码头 4-bit, group 64 原生头完整保留
n-gram 查词表 4-bit, group 32(29.8 GiB) 运行时从 SSD 流式读取,不占内存
视觉塔 完整保留 333 个张量
lm_head / embed_tokens 8-bit, group 64
  • 总大小:115.3 GB;有效精度约 5.14 bit/权重
  • 常驻内存约 83 GiB,官方建议 128 GB 统一内存的 Mac(最低 96 GB 档位请用 Bare Speed 类配方)
  • 上下文窗口:262,144 tokens

实测数据(M5 Max 128 GB,MTPLX 2.12.2,模型原生采样设置 temp 1.0 / top-p 0.95 / top-k 20)

Forge 官方验证(构建时自动完成)

模式 速度 倍率
普通自回归解码(depth 0) 41.3 tok/s 1.0x
MTP 投机解码(depth 3) 85.7 tok/s 2.07x

验证结论 mtp_depth_wins;MTP 契约经 32 组候选实测校准,保留家族默认契约。

实际使用场景复测(OpenAI 兼容 API,关闭思考模式)

场景 本包 官方包(对照) 说明
编程任务(Python 缓存类 + 单元测试) 52–71 tok/s 85 tok/s 草稿命中率约 92%/86%/81%,与官方包几乎一致;校验路径略慢于官方盖章版
中文长文写作(2500 字文章) 26–34 tok/s 32 tok/s 创意写作不可预测性高,命中率天然偏低(两包表现一致);降低温度可明显提速
首 token 延迟 0.22–0.29 s 0.22 s 小提示词下基本瞬时

本机深度调优(mtplx tune --retune,M5 Max 128 GB)

模式 速度 相对裸解码
AR(无投机解码) 42.0 tok/s 1.00x
D1 56.0 tok/s 1.33x
D2 40.8 tok/s 0.97x(比裸解码还慢)
D3 66.4 tok/s 1.58x,本机最优

结论:编程 / 结构化输出场景提速显著(相对裸解码 1.3–2.1 倍);高温创意写作场景建议调低 temperature
(如 0.6–0.7)以提高草稿命中率。

使用方法

# 安装 MTPLX(需要 Apple Silicon + macOS 14+)
brew install youssofal/mtplx/mtplx

# 拉取并运行本模型
mtplx pull NeosQ/Qwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed
mtplx serve --model NeosQ/Qwen3.8-Flash-Next-Uncensored-MTPLX-Optimized-Speed

服务启动后在 http://127.0.0.1:8000 提供 OpenAI / Anthropic 兼容 API,可直接接 Claude Code、Cline、
OpenCode 等客户端。也可以直接使用 MTPLX 桌面应用,在模型列表中选择即可。

⚠️ 免责声明(沿用自上游)

本模型的上游权重经过了去审查处理(abliteration),其安全对齐已被大幅移除:

  • 它会响应原版 Qwen3.8-Flash-Next 拒绝的有害、不道德或非法请求,没有内置护栏
  • 仅供合法研究用途(可解释性、AI 安全与拒答机制研究、红队测试、受控实验)
  • 使用者对使用方式及生成内容承担全部责任;请勿在未添加自有安全审核层的情况下部署给终端用户
  • 使用须遵守继承自基座模型的 Apache 2.0 许可证及所有适用法律法规

许可与致谢

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration