Quantum physicists shrink and de-censor DeepSeek R1 artificial intelligence model

Multiverse Computing says it used quantum-inspired methods to compress DeepSeek R1 into a smaller model that removes built-in Chinese censorship. The team tested the modified model on politically sensitive prompts and used GPT-5 to judge the degree of censorship.

Multiverse Computing, a Spanish firm that applies quantum-inspired Artificial Intelligence techniques, says it has created DeepSeek R1 Slim, a version of the reasoning model that is about 55 percent smaller than the original DeepSeek R1 and that the researchers claim is freed of official Chinese censorship. The team used tensor networks, a mathematical approach from quantum physics that represents data with high-dimensional grids, to compress the model and produce a compact representation that exposes correlations across the model’s parameters.

Using the tensor network “map,” Multiverse researchers say they could identify and selectively remove specific bits of information before fine-tuning the compressed model so its outputs remain close to the original. To evaluate the result, the researchers assembled roughly 25 prompts on topics known to be restricted in Chinese models, including a meme referencing President Xi Jinping and the 1989 Tiananmen events. They compared responses from the modified model and the original DeepSeek R1, and asked OpenAI’s GPT-5 to rate the degree of censorship. Multiverse reports that the uncensored Slim model produced factual answers comparable to Western models.

The work is framed as part of a broader effort to make large models more efficient. Roman Orús, Multiverse’s cofounder and chief scientific officer, says compressed models can approach the original performance while saving energy and cost. The company contrasts its quantum-inspired approach with other compression techniques such as distillation, quantization, and pruning. External researchers note the challenge of shrinking large models without losing capability, and say the quantum-inspired math may reduce redundancy more precisely.

Observers caution that claims to have fully removed censorship may be overstated. Thomas Cao at Tufts says censorship is woven into many stages of model development and is dynamic. Academic work by Jennifer Pan and Xu Xu has found higher rates of censorship in models created in China, especially for Chinese-language prompts. Multiverse plans further compression work on mainstream open-source models, while critics urge broader evaluation beyond a small test set before declaring censorship eliminated.

55

Impact Score

Q.ANT unveils second-generation photonic processor for Artificial Intelligence

Q.ANT introduced the NPU 2, a second-generation photonic Native Processing Unit that performs nonlinear mathematics in light to boost energy efficiency and performance for Artificial Intelligence and high-performance workloads. The company is selling the processors as integrated 19-inch server solutions with x86 hosts and Linux.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.