Can Quantum-Inspired AI Compete With Today's Large Language Models?

Forbes

19-04-2025

Quantum AI
getty
As large language models (LLMs) continue their rapid evolution and domination of the generative AI landscape, a quieter evolution is unfolding at the edge of two emerging domains: quantum computing and artificial intelligence. A growing cohort of researchers and companies are now exploring whether principles of quantum computing can address some of the limitations facing today's AI infrastructure, particularly around scalability, efficiency, and reasoning complexity.
One of the more notable developments comes from Dynex, a Liechtenstein-based firm that recently unveiled its Quantum Diffusion Large Language Model (qdLLM) as a finalist in the SXSW 2025 innovation awards. The company claims that its qdLLM can generate genAI outputs faster and more efficiently than traditional transformer-based systems that depend on current technology infrastructure.
How does this compare to other emerging approaches, and what does it mean for the broader future of AI?
At its core, quantum computing differs from classical computing by leveraging quantum bits (qubits), which can exist in multiple states simultaneously thanks to quantum superposition. This allows quantum computers to evaluate a vast number of potential solutions in parallel, potentially offering advantages in tasks that require large-scale optimization, simulation, or pattern recognition.
In the context of AI, researchers have explored how quantum properties might improve tasks like natural language processing, machine learning optimization, and model training efficiency. However, most of these efforts remain at an early stage. For example, IBM and MIT have studied how hybrid quantum-classical models could reduce training time for specific deep learning tasks, while startups like Zapata AI are experimenting with quantum-enhanced models for sentiment analysis and forecasting.
Against this backdrop, Dynex's approach introduces a new architecture that uses quantum-inspired algorithms to run LLMs more efficiently via decentralized hardware.
Unlike transformer-based models that generate responses one token at a time using autoregressive techniques, Dynex's qdLLM is built on a diffusion model that creates output tokens in parallel. According to Dynex, this approach is more computationally efficient and yields better contextual consistency.
'Traditional models like GPT-4 or DeepSeek work sequentially,word after word,' says Daniela Herrmann, Co-founder and Global Head of Leadership at Dynex. 'qdLLM works in parallel. It thinks more like the human brain,processing patterns all at once. That's the power of quantum.'
Several academic projects, including those at Stanford and Google DeepMind, as well as initiatives out of the major AI technology providers have recently begun exploring diffusion-based transformers.
Dynex further differentiates itself by integrating quantum annealing, a form of quantum optimization, to improve token selection during text generation. The company claims that this improves coherence and reduces computational overhead compared to conventional LLMs.
One unique feature of Dynex's model is its reliance on a decentralized network of GPUs that emulate quantum behavior, rather than requiring access to actual quantum hardware. This design allows the system to scale to what Dynex describes as up to one million algorithmic qubits.
Herrmann explains, 'Any quantum algorithm, for example qdLLM, is being computed on a decentralised network of GPUs which are efficiently emulating the quantum calculations.'
This type of emulation draws some parallels to the work of TensorFlow Quantum (by Google and X), which also simulates quantum circuits on classical hardware to prototype algorithms. Similarly, a number of technology startups and vendors are developing platforms to simulate quantum logic at scale before physical hardware is ready.
In addition to software, Dynex plans to introduce its own neuromorphic quantum chip, named Apollo, by 2025. Unlike superconducting quantum chips that require cryogenic cooling, Apollo is designed to run at room temperature, supporting integration into edge devices.
'Using neuromorphic circuits allows Dynex to emulate quantum computing on a large scale, of up to 1 Million algorithmic qubits,' Herrmann explains. 'Dynex will start producing actual quantum chips, which are also based on the neuromorphic paradigm.'
Dynex states that qdLLM achieves 90% smaller model sizes, is 10x faster, and uses only 10% of the GPU resources typically required for equivalent tasks. These are significant claims, particularly given increasing scrutiny on AI's energy usage.
'The efficiency and parallelism of the quantum algorithm reduces the energy consumption, because it is 10x faster and requires only 10% of the number of GPUs,' says Herrmann.
While independent verification is still needed, Dynex's approach echoes efforts by Cerebras Systems, which has created wafer-scale chips that use less energy per training task. Another example is Graphcore, whose intelligence processing units (IPUs) aim to reduce the energy footprint of AI workloads through specialized parallel architecture.
Dynex reports that qdLLM is performing strongly against leading models,including ChatGPT and Grok, on benchmarks requiring strong reasoning. Though public benchmark data has not yet been released, the company states it will publish comparative studies as it moves closer to a 2025 market launch. Until peer-reviewed benchmarks are available, Dynex's performance assertions remain anecdotal but intriguing.
'We are publishing qdLLM benchmarks regularly and have demonstrated that certain questions, which require strong reasoning, cannot be correctly answered by ChatGPT, Grok or DeepSeek,' Herrmann notes.
In the long term, Dynex sees quantum computing becoming central to the AI field.
'We think that quantum will be dominating AI in the next 5 years,' Herrmann says.
That projection remains speculative, though it's not without precedent. Analysts from McKinsey, BCG, and Gartner all note that quantum computing could dramatically improve optimization and simulation tasks, but likely not until beyond 2030 for most use cases. A more measured view suggests that quantum-AI hybrids will emerge first in niche applications, such as drug discovery, financial risk modeling, or cybersecurity.
For now, Dynex stands among a growing field of players experimenting with quantum-enhanced or quantum-inspired AI methods. Whether their decentralized, diffusion-based qdLLM can scale beyond benchmarks remains to be seen, but its arrival signals that the search for new foundations in AI is far from over.