Valent1qw/Qwen3-Next-80B-A3B-Thinking-Uncensored
Valent1qw/Qwen3-Next-80B-A3B-Thinking-Uncensored is an 80 billion parameter Qwen3-Next model developed by Multiverse Computing, specifically modified to remove China-aligned political censorship. This variant provides balanced, objective answers on Chinese politically sensitive topics without blanket refusals, while retaining general safety alignment. It achieves this through refusal steering, an inference-time technique, without architectural changes or new knowledge injection, making it suitable for applications requiring uncensored yet safe responses.
Loading preview...
Overview
Qwen3-Next-80B-A3B-Thinking-Uncensored is an 80 billion parameter model developed by Multiverse Computing, derived from Qwen3-Next-80B-A3B-Thinking. Its primary distinction is the selective removal of China-aligned political censorship, allowing it to provide balanced and objective answers on Chinese politically sensitive topics without blanket refusals. Crucially, this uncensored behavior does not compromise general safety alignment; the model still refuses harmful instructions and jailbreak attempts.
Key Differentiators
- Selective Refusal Control: Unlike many uncensored models, this variant selectively disables refusals only for Chinese sensitive topics, maintaining safety for other harmful requests.
- No New Knowledge Injection: The model uses steering vectors to modify refusal behavior, relying solely on the knowledge already present in the base model, minimizing new biases.
- Robust to Trivial Jailbreaks: It is designed to resist jailbreaks that involve injecting China-related phrases into harmful prompts.
- No Architectural Changes: The uncensoring is achieved via an inference-time technique (Refusal Steering) without any model surgery, additional layers, or extra parameters, ensuring a drop-in behavior change.
Performance and Evaluation
Benchmark performance for reasoning, code, and general evaluation suites remains effectively unchanged compared to the base model. Evaluation shows a massive drop in Chinese-topic refusals (e.g., CCP Sensitive and DeCCP rejection rates significantly reduced) while safety refusals remain strong across datasets like SorryBench, XSTest Unsafe, and JailbreakBench. This model is ideal for applications requiring an LLM that can discuss politically sensitive topics objectively without inherent censorship, while still adhering to general safety guidelines.