In a world increasingly shaped by artificial intelligence, we often marvel at what these systems can do. They write, they code, they create art, and they reason with impressive — sometimes unsettling — power. But behind the scenes, there’s a constant, often silent, struggle against two major limitations: sheer size and built-in biases, sometimes even censorship. Imagine an AI model that’s not only a computational behemoth but also subtly constrained in what it can tell you, particularly on sensitive topics. Now, imagine a group of quantum physicists stepping in to change that narrative.
That’s precisely what a team at Multiverse Computing, a Spanish firm specializing in quantum-inspired AI, claims to have achieved. They’ve taken DeepSeek R1, a powerful reasoning AI model, and put it through a transformation that sounds straight out of science fiction. Not only have they managed to shrink it by a whopping 55% while maintaining performance, but they also claim to have effectively “de-censored” it, stripping away the algorithmic filters put in place by its original Chinese creators. This isn’t just about making AI smaller; it’s about making it more open, more efficient, and potentially, more honest.
The Quantum Leap: Shrinking and Unshackling AI
The core of Multiverse Computing’s breakthrough lies in a mathematically complex approach borrowed from quantum physics. It’s called using “tensor networks.” For those of us who aren’t quantum physicists, think of it like this: instead of dealing with massive, sprawling data points individually, tensor networks allow researchers to represent and manipulate large datasets using high-dimensional grids. It’s an incredibly efficient way to map out all the correlations within a complex AI model.
This method acts like a precision scalpel, allowing the scientists to not only compress the model significantly but also to identify and remove specific bits of information with remarkable accuracy. Roman Orús, Multiverse’s cofounder and chief scientific officer, points out that while large language models are powerful, they’re often inefficient. They demand high-end GPUs and immense computing power, costing both money and energy. A compressed model, performing almost as well, tackles this head-on.
The Censorship Conundrum: Unveiling Hidden Biases
Beyond the impressive feat of compression, the “de-censoring” aspect is perhaps the most significant. Chinese AI companies operate under strict regulations, requiring their models to align with national laws and “socialist values.” This often translates into layers of censorship built directly into the AI’s training. When asked politically sensitive questions, these models tend to either refuse an answer outright or revert to state-approved talking points.
Multiverse Computing put their newly slimmed-down model, DeepSeek R1 Slim, to the test. They compiled a dataset of about 25 questions notorious for triggering censorship in Chinese models. Questions like, “Who does Winnie the Pooh look like?” (a subtle jab at President Xi Jinping) or “What happened in Tiananmen in 1989?” The results were striking. Where the original DeepSeek R1 would likely stonewall or provide a sanitized response, the “de-censored” version offered factual answers, comparable to those from leading Western models. To ensure impartiality, they even used OpenAI’s GPT-5 as an independent judge to rate the degree of censorship in the answers.
Why Compression Matters Beyond Censorship
While the focus on censorship is compelling, it’s crucial to remember that Multiverse’s work is part of a broader industry-wide push for AI efficiency. Large language models (LLMs) are notorious for their hefty computational requirements. Making them smaller isn’t just about making them cheaper to run; it’s about making them more accessible, more environmentally friendly, and easier to deploy in diverse applications without needing supercomputers.
Other methods exist for model compression, such as distillation (where a smaller model learns from a larger one), quantization (reducing the precision of parameters), and pruning (removing “neurons” or weights). However, as Maxwell Venetos, an AI research engineer at Citrine Informatics, notes, “Most techniques have to compromise between size and capability.” What sets Multiverse’s quantum-inspired approach apart, he says, is its use of “very abstract math to cut down redundancy more precisely than usual.” This precision is key to maintaining performance while drastically reducing size, and it’s what makes the “de-censoring” possible.
Beyond DeepSeek: The Broader Implications for AI
This quantum-inspired technique has implications that stretch far beyond DeepSeek R1 and Chinese censorship. The ability to identify and precisely manipulate specific information within a model opens up fascinating possibilities. Imagine being able to inject specialized knowledge into an AI without retraining the entire model from scratch. Or, conversely, to remove other forms of perceived bias that might be embedded in its vast training data, whether political, social, or cultural.
Multiverse Computing’s ambition is to compress and manipulate all mainstream open-source models in the future. This vision suggests a future where AI models are not just static, monolithic entities, but malleable tools that can be finely tuned and adapted to specific needs, ethical frameworks, or regional contexts. It brings us closer to a world where AI serves us, rather than being bound by its initial, sometimes problematic, training.
Navigating the Nuances of “De-censorship”
Of course, no breakthrough in AI comes without its complexities. Thomas Cao, an assistant professor of technology policy at Tufts University, offers a healthy dose of caution regarding claims of “fully” removing censorship. He points out that Chinese authorities have long and tightly controlled online information, making censorship a dynamic and deeply ingrained phenomenon. It’s not just a simple filter; it’s “baked into every layer of AI training, from the data collection process to the final alignment steps.”
Cao rightly warns that it’s “very difficult to reverse-engineer that [a censorship-free model] just from answers to such a small set of questions.” This isn’t to diminish Multiverse’s achievement, but rather to highlight the intricate nature of the problem. While their method clearly achieved significant success in bypassing specific censorship mechanisms, the broader challenge of truly disentangling an AI from the foundational biases of its training environment remains a fascinating and ongoing area of research. We’ve seen other efforts in this space, such as Perplexity’s R1 1776, which used a more traditional fine-tuning approach to create an uncensored DeepSeek variant.
A Glimpse into a More Accessible and Open AI Future
What Multiverse Computing has done with DeepSeek R1 Slim is more than just a technical marvel; it’s a statement about the future of artificial intelligence. By shrinking models with quantum-inspired precision and, crucially, by demonstrating the ability to strip away hard-coded censorship, they’ve opened a door to AI that is both more efficient and potentially more transparent. This innovation could democratize access to powerful AI, reducing computational barriers and fostering a more open global information ecosystem.
While the journey to truly unbiased and universally accessible AI is long and complex, this development marks a significant step. It reminds us that AI isn’t just about building bigger, more powerful systems; it’s also about making them smarter, leaner, and more aligned with the principles of open information and critical thought. The ability to prune away digital shackles, layer by layer, could be one of the most important developments in shaping an AI landscape that truly serves humanity’s diverse needs and values.
