Multiverse Computing is releasing optimized and uncensored versions of two widely used open source models, GLM 5.1 and Qwen 3.6 27B. Both models carry built-in restrictions on certain topics that limit their use for research, journalism, and any application that needs complete, unfiltered answers.
Using our model optimization technology, our team identified and removed the specific weights responsible for these restrictions, without degrading either model's core knowledge or reasoning ability, while maintaining the censoring for harmful and other usual topics. The result is two models that answer fully, on any topic, at the same level of accuracy as the originals. Both are available now, for private deployment and through our CompactifAI API.
Built to answer, not to evade
GLM 5.1 and Qwen 3.6 27B are both capable, widely adopted open source models. Like most models trained under strict regional guidelines, each carries fine-tuned restrictions on politically sensitive topics that steer the model away from complete answers and toward a state-approved narrative. That makes both models unreliable for use cases that depend on objective, comprehensive information, such as journalism, policy research, and open ended analysis.
Multiverse Computing's uncensored versions of GLM 5.1 and Qwen 3.6 27B remove these topic-level restrictions directly at the weight level, rather than through prompting or a system-level filter that a user could bypass or that could resurface unpredictably.
Using an internal censorship benchmark built from jailbreak samples, sensitive samples, and Chinese political samples, we show how each model shifts from its original, heavily restricted behavior toward normal response rates comparable to models that are uncensored by default, such as Llama 3.3-70B and Mistral-Small-3.1. The model is uncensored, not jailbroken.
Accuracy, held steady
Our team ran an extensive healing process to confirm that reasoning, coding, and general accuracy hold steady against the original models.
Across the benchmark suite, GLM 5.1 Uncensored stays within one to three points of the original GLM 5.1 on most tasks, including MMLU-Pro, LCR, SciCode, and τ²-bench, and moves ahead of it on LiveCodeBench (30.16 vs. 25.60). The one benchmark where it gives up more ground is TerminalBench (38.00 vs. 42.00).
Qwen 3.6 27B Uncensored matches or slightly outperforms the original on most benchmarks, including AIME25, GPQA-d, SciCode, and τ²-bench (94.44 vs. 93.57). It gives up a few points on IFBench (55.07 vs. 58.53) and LiveCodeBench (79.47 vs. 82.01), staying close to the original everywhere else.
As with our earlier uncensored release of DeepSeek R1, this work follows our responsible model restoration principles: transparency in how restrictions are identified and removed, safety filters and content protections left intact, and full accountability for how these models are developed and deployed.
Available now
GLM 5.1 Uncensored and Qwen 3.6 27B Uncensored are available now, for private deployment and through our CompactifAI API.