На Hugging Face выложили почти полностью разцензуренную Qwen3.5-9B. 😳 Модель называется DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED. Разработчики сначала дообучили её на 250 тысячах примеров по математике и естественным наукам, сгенерированных DeepSeek-V4-Pro, а затем модифицировали, чтобы она реже отказывалась выполнять запросы. Результат: в тесте автора из 100 запросов оригинальная модель отказала 100 раз, а модифицированная - всего 3. Причём это всего 9B-параметров. Версия с квантованием Q4 весит около 5,6 ГБ, так что запускать её можно локально на обычном компьютере. https://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
Research1 min read
An almost completely uncensored Qwen3.5-9B was posted on Hugging Face. 😳
An almost completely uncensored Qwen3.5-9B was posted on Hugging Face. 😳 The model is called DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED. The developers first trained it on 250 thousand examples in mathematics and natural sciences generated by DeepSeek-V4-Pro, and then modified it…
Telegram — Вайб-кодингBy: Вайб-кодингNovember 8, 2026 (1 hour ago)
Available in 4 languagesAI Summary & Key Takeaways
- An almost completely uncensored Qwen3.5-9B was posted on Hugging Face
- 😳 The model is called DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED
- The developers first trained it on 250 thousand examples in mathematics and natural sciences generated by DeepSeek-V4-Pro, and then modified it…
Research
KZ·AI D100Инсайт дня
An almost completely uncensored Qwen3.5-9B was posted on Hugging Face. 😳
Full text available in Russian
Related stories
Investors push for Anthropic valuation cut ahead of IPO
✔️ Investors are pushing for Anthropic's valuation to be lowered ahead of its IPO Anthropic's plans to hold a roadshow and go public in November are in jeopardy. Potential investors believe the company's $1.5 trillion valuation is too high and are demanding a discount. Skepticism on Wall Street...
Telegram — AI Machine Learning Big Data
Andrej Karpathy built a curated video platform in a weekend using Claude and 20 years of Hacker News
OpenAI co-founder Andrej Karpathy used Claude's Opus 5.5 to filter 43 million Hacker News comments into 6,454 curated videos. The result beats YouTube's own recommendations, he says. The article Andrej Karpathy built a curated video platform in a weekend using Claude and 20…
The Decoder
AI News Digest for November 8, 2026
The day proved telling: independent evaluators from METR note that Anthropic is accelerating AI development twice as fast as a year ago, but some figures in the report are obscured. Washington meanwhile approves...
SMNTCN (Semanticon) — глобальные новости ИИ
The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts
Turing Award winner Richard Sutton reaffirms that raw compute alone isn't enough for intelligence and that fundamentally new learning methods are needed. Ex-OpenAI researcher Tworek agrees, saying blind transformer scaling "has reward hacked brains of a whole generation of…
The Decoder