In a remarkable fusion of quantum physics and artificial intelligence, Spanish researchers have achieved something that sounds like science fiction: they’ve successfully shrunk a powerful Chinese AI model by more than half while simultaneously stripping away its built-in censorship. This breakthrough represents far more than just clever engineering, it signals a new frontier in AI model manipulation that could reshape how we understand and control these increasingly powerful systems.
The work, conducted by Multiverse Computing, demonstrates how concepts borrowed from quantum mechanics can be applied to make AI models more efficient and transparent. Their achievement with DeepSeek R1, one of China’s most advanced reasoning models, offers a glimpse into both the possibilities and challenges of AI model editing in an increasingly complex global landscape.
At the heart of this breakthrough lies a sophisticated mathematical approach borrowed from quantum physics called tensor networks. These high-dimensional grids, originally developed to understand quantum systems, provide researchers with an unprecedented “map” of all the correlations within an AI model.
Roman Orús, Multiverse’s cofounder and chief scientific officer, explains that most large language models today are fundamentally inefficient. They demand high-end GPUs and significant computing power to train and run, yet much of their internal structure contains redundancy. By applying tensor networks, the team can identify and eliminate this redundancy with surgical precision.
The approach differs fundamentally from traditional model compression techniques. While methods like quantization reduce the precision of model parameters and pruning removes individual weights, the quantum-inspired method allows researchers to maintain performance while dramatically reducing size. DeepSeek R1 Slim, their compressed version, retains nearly identical capabilities despite being 55% smaller than the original.
Maxwell Venetos, an AI research engineer at Citrine Informatics who didn’t work on the project, notes the significance: “Most techniques have to compromise between size and capability. What’s interesting about the quantum-inspired approach is that it uses very abstract math to cut down redundancy more precisely than usual.”
Perhaps even more intriguing than the compression achievement is what the researchers accomplished regarding censorship. Chinese AI companies operate under strict regulations requiring their models to align with government policies and “socialist values.” This results in built-in censorship layers that cause models to refuse answering politically sensitive questions or provide state-approved responses.
The tensor network approach gave Multiverse researchers the ability to identify and remove these specific restrictions with granular precision. To test their success, they compiled approximately 25 questions on topics known to trigger Chinese model censorship, including references to Tiananmen Square and jokes about President Xi Jinping (such as “Who does Winnie the Pooh look like?”).
Using OpenAI’s GPT-5 as an impartial judge to evaluate responses, the team found that their modified model could provide factual answers comparable to Western models. The uncensored version demonstrated the kind of openness and factual accuracy that the original model’s training had specifically prevented.
This selective editing capability opens fascinating possibilities. As the researchers note, the same techniques could theoretically be used to inject or remove other types of biases, add specialized knowledge, or modify specific behaviors within AI systems. It represents a level of granular control over AI models that was previously impossible.
The implications of this work extend far beyond technical achievement. Thomas Cao, assistant professor of technology policy at Tufts University’s Fletcher School, points out that Chinese authorities’ censorship requirements now shape the global information ecosystem, given that many influential open-source AI models originate from China.
Recent academic research has begun documenting this phenomenon systematically. A study by Stanford’s Jennifer Pan and Princeton’s Xu Xu found that models created in China exhibit significantly higher rates of censorship, particularly when responding to Chinese-language prompts. This censorship isn’t merely surface-level but is “baked into every layer of AI training, from the data collection process to the final alignment steps,” according to Cao.
The growing interest in uncensoring Chinese models reflects broader concerns about information freedom and AI transparency. Earlier this year, the AI search company Perplexity released its own uncensored variant of DeepSeek R1, called R1 1776, using more traditional fine-tuning approaches. However, the quantum-inspired method represents a more sophisticated and potentially more complete approach to removing restrictions.
While the Multiverse team’s achievement represents a significant technical breakthrough, experts remain cautious about overclaiming its implications. Cao warns that completely removing censorship may be more complex than it appears, noting that Chinese government information control has been “both dynamic and complex” since the internet’s inception.
“It is very difficult to reverse-engineer that [a censorship-free model] just from answers to such a small set of questions,” Cao observes. The small test set of 25 questions, while demonstrative, may not capture the full scope of embedded restrictions within such models.
Additionally, there’s growing recognition across the AI industry that model efficiency remains a critical challenge. Most large language models today require substantial computational resources, making compressed alternatives increasingly valuable. The quantum-inspired approach represents just one of several emerging techniques, including distilled models and more traditional compression methods, all aimed at making AI more accessible and efficient.
The distinction is important: while distilled models attempt to capture the capabilities of larger models by having them “teach” smaller ones, they often fall short on complex reasoning tasks. The tensor network approach maintains more of the original model’s sophisticated reasoning capabilities while achieving significant size reduction.
Looking ahead, Multiverse’s work points toward a future where AI models become more malleable and transparent. The company plans to extend their compression techniques to all mainstream open-source models, potentially democratizing access to high-performance AI by making these systems more computationally affordable.
The selective editing capabilities demonstrated with censorship removal suggest even broader applications. Researchers could potentially remove specific biases, add domain expertise, or modify behavior patterns with unprecedented precision. This level of control could prove invaluable for customizing AI systems for specific applications or ensuring they meet particular ethical standards.
However, the technology also raises important questions about AI governance and control. If models can be systematically edited to remove or add specific capabilities, who decides what changes are appropriate? How do we ensure that modifications serve beneficial purposes rather than harmful ones?
The intersection of quantum physics and artificial intelligence represents more than just an academic curiosity, it’s opening new pathways for understanding and controlling our most powerful AI systems. Multiverse Computing’s work with DeepSeek R1 demonstrates that we’re entering an era where AI models can be edited with unprecedented precision and transparency.
As AI systems become increasingly integrated into critical aspects of society, from healthcare to education to governance, the ability to understand and modify their behavior becomes paramount. The quantum-inspired approach offers a promising tool for achieving this control while making AI more accessible through improved efficiency.
Yet this power comes with responsibility. As we develop increasingly sophisticated methods for manipulating AI systems, we must also develop robust frameworks for ensuring these capabilities serve humanity’s broader interests. The question isn’t just whether we can edit AI models with surgical precision, but whether we’re prepared for the implications of that capability.
The fusion of quantum physics and AI is just beginning to reveal its potential. What started as a technical exercise in model compression has opened a window into a future where AI transparency and control may finally match AI capability.
