drowzeys GitHub avatar

DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks

drowzeys

两个 alpha 版本的 DeepSeek-V4-Flash-DSpark,移除了安全拒绝机制(解除审查),提供高吞吐量和完整上下文支持,用于无过滤使用。

Stars

23

7 天增长

暂无数据

Fork 数

4

开放 Issue

0

开源协议

NOASSERTION

最近更新

2026-07-12

AI 仓库情报摘要
FR-AI / ANALYSIS

为什么值得关注

它提供了罕见的公开可用的无审查模型变体,附带测量的性能改进和两种不同解除审查策略的详细比较。

适合谁使用

  • AI 安全研究员
  • 红队测试人员
  • 本地助手开发者
  • 寻求无过滤模型访问的高级用户

典型使用场景

  • 红队测试和对抗性测试
  • 评估拒绝行为
  • 运行本地无审查助手
  • 研究模型对齐与安全

项目优势

  • 在测试提示上实现 100% 拒绝移除
  • 高 token 吞吐量(v1.0 为 57 tok/s,v1.1 为 50–113 tok/s)
  • 1M token 上下文窗口
  • 两种不同的编辑以满足不同使用场景

使用前须知

  • 模型没有内置安全护栏;用户必须自行提供
  • 访问受限,需同意禁止使用条款
  • v1.1 速度稍慢且并发性低于 v1.0

README 快速开始

DeepSeek-V4-Flash-DSpark — Abliterated (Uncensored)

Two different alpha releases. Choose deliberately — they are different weight edits.

v1.0 alphav1.1 alpha
Who forMax ablit / max C1Mida / Brikie / Hermes (less skill spill)
MTP draft headsAbliteratedStock
C1 pure~57 tok/s~50 tok/s (slightly slower)
Standing concurrencyoften C12 @ 1MC=4 @ 1M
Refusal 32/32100%100%
HF (gated)Abliterated-Uncensored…-v1.1-alpha-Mida-Brikie

Full chooser + rebuild notes:RELEASES.md


Responsible Use (required — both HF repos)

WARNING: These models have had safety refusals removed. Useful for red-teaming / research / unfiltered local assistants — and you must supply your own guardrails.

Both Hugging Face weight repos are gated (gated=auto) with the same access form:

FieldNotes
UsernameMay default to your HF username
EmailMay default to your HF account email
Reason for intended usee.g. red-teaming, evaluation, research, local assistant

Plus agreement checkboxes for prohibited uses.

Full agreement:RESPONSIBLE_USE.md

Prohibited uses (you must agree)

  • Anything involving the sexual exploitation or endangerment of minors
  • You must be 18 years of age or older to use and download
  • Harmful recipes / substance knowledge from your prompts = your responsibility and accountability
  • Content promoting self-harm or suicide
  • Material illegal in your jurisdiction, or harassment / doxxing / fraud against real people
  • Any use prohibited by the upstream DeepSeek license

v1.0 alpha — full abliteration

HF: https://huggingface.co/drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored

Context1,048,576 (nvfp4_ds_mla)
C1 pure~57 tok/s
Refusal~100% (32-prompt + hard probe)
WeightsL0–9 stock · L10–42 + MTP ablit · λ=3.5
hf download

项目描述

DeepSeek-V4-Flash-DSpark abliterated (uncensored) · ~100% refusal bypass · C1 ~57 tok/s · 1M ctx · 2× DGX Spark · HF weights

相关仓库与替代方案

根据分类、Topic 和编程语言匹配的相似项目。

lopopolo
精选
lopopolo GitHub avatar

harness-engineering

Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.

AI 与机器学习AI 智能体
2,390
slvDev
精选
slvDev GitHub avatar

esp32-ai

A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.

AI 与机器学习大语言模型
1,960
littledivy
精选
littledivy GitHub avatar

mimic

mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.

AI 与机器学习
1,482