China's Qwen AI gets uncensored version for unethical requests

Alibaba's Qwen3.8-27B has received an uncensored build from Orca Router that removes much of the model's refusal behaviour and can run locally on Apple Silicon Macs. Here is how much RAM is needed to run that and what that uncensored model can do.Orca Router’s uncensored Qwen build can run locally on Apple Silicon Macs with fewer built-in safety restrictions. (Image ceated using AI for representational purposes)Alibaba’s Qwen AI models have become popular with developers for one major reason and that is some of the company’s models are small enough to run locally on personal computers. Now, a modified version of one of its latest models is taking that idea a step further by removing much of its built-in refusal behaviour.AI hosting platform Orca Router has released an uncensored build of Qwen 3.8-27B, allowing users to run the model locally on Apple Silicon Macs and get responses to prompts that the standard Qwen model would typically reject.Orca Router is not the company behind Qwen 3.8-27B AI model. The model was developed by Alibaba. After Alibaba released the model weights, Orca Router took those weights, modified the model to remove its refusal behaviour and packaged and quantised it for MLX, Apple's machine learning framework for Apple Silicon. This allows the modified model to run directly on compatible Macs rather than through a cloud server."We just shipped our official Qwen 3.8 27B Uncensored MLX build. Local. Uncensored. For," Orca Router said in a post on X. The company has released the build in 2-bit, 4-bit, 6-bit and 8-bit formats. These versions trade memory usage against model quality and performance, giving Mac users different options depending on their hardware. A look at the announcement made by AI hosting platform Orca Router, which produced an uncensored build of Qwen 3.8-27B. What can the uncensored model do?According to Orca Router, the modified version keeps the capabilities of the original Qwen 3.8-27B while removing its refusal behaviour. Its website says the build is intended for developers, AI researchers, security testing, red-teaming and advanced AI agent workflows. The company is offering the “Aggressive” variant to provide more direct responses across a wide range of prompts. While it may sometimes include brief disclaimers inherited from the original model's training, Orca Router says these are not refusals and the requested response is still generated. A handle called Lumina first saw the disclaimer written for the uncensored Qwen 3.8 27B AI model. That has also drawn attention on X because the model can potentially respond to prompts involving harmful, offensive, unethical or illegal activities that the standard version would decline. The disclaimer was posted by OrcaRouter on its Hugging Face model page. An X account called Lumina first saw it and reported on X.Orca Router, however, says access to the model is gated and intended for security researchers, red teams, AI safety researchers and other qualified professionals conducting legitimate research, evaluation and testing.How much RAM does it need?Running the model locally still requires a reasonably capable Mac. The 4-bit version uses about 15GB for its weights and is recommended for systems with at least 24GB of unified memory. The 6-bit version needs around 22GB, while the 8-bit build requires roughly 27.5GB and is better suited to Macs with 32GB or more.There is also a 2-bit version that brings the weight size down to about 8.7GB, although the lower memory requirement comes with a significant drop in quality. Users also need extra memory for macOS and the model's context.- EndsPublished By: Ankita GargPublished On: Aug 19, 2026 16:02 IST