en.Wedoany.com Reported - Photonics computing company Q.ANT successfully demonstrated diffusion models and recurrent neural networks on its second-generation native processing unit at the ISC High Performance 2026 conference in Hamburg, proving that its photonic architecture can support modern AI workloads including generative image synthesis and time-series sequence prediction.

This demonstration by Q.ANT builds on ongoing ecosystem progress. Previously, independent developer Daisytuner achieved the compilation and deployment of an object detection model from PyTorch onto Q.ANT's photonic processor, marking the first successful compilation of an AI model from a standard machine learning framework onto photonic hardware.

Q.ANT stated that its photonic native processing system has evolved from the basic algorithm stage to commercial application. At the photonic circuit level, Q.ANT hardware targets 30 times the energy efficiency of traditional processors when performing equivalent matrix operations.

Dr. Michael Förtsch, founder and CEO of Q.ANT, noted that the photonic architecture changes the energy consumption calculation method for AI infrastructure. Performing computations with light rather than transistors reduces energy consumption at the source, which is a bottleneck the AI industry must overcome. He emphasized that recent generative AI demonstrations show photonic hardware can handle the mathematical load of the most demanding modern AI workloads.

To showcase generative AI capabilities, Q.ANT hardware ran a diffusion model for image-to-image synthesis, a workload defined by iterative, parallelized matrix operations. Q.ANT stated this is the first time a diffusion model of this complexity has been run on photonic hardware. Diffusion models generate images through repeated forward propagation of deep neural networks in dense matrix operations. Q.ANT's photonic processor uses light instead of transistors to execute the main computational layers, thereby entering the realm of linear arithmetic at the core of modern AI applications.

Professor Björn Ommer, head of the Computer Vision and Learning Group at Ludwig Maximilian University of Munich, stated that diffusion models are widely used and computationally intensive methods in modern generative AI. If photonic hardware can efficiently and reliably execute such workloads, it indicates that alternative computing substrates could play a significant role in the future of generative AI.

Q.ANT also executed the TiRex time-series prediction model developed by NXAI, an Austrian frontier AI lab, which is based on an extended long short-term memory architecture. Lukas Fischer, Head of Applied Research at NXAI, stated that TiRex aims for a balance between performance and power consumption, and that the xLSTM architecture on photonic systems could redefine what energy-efficient AI means. Unlike Transformer-based models, xLSTM is a recurrent neural network used to identify patterns in sequence data and predict future values over long time horizons. NXAI's commercial TiRex model uses production-tuned weights, targeting applications in financial market analysis, supply chain optimization, weather forecasting, and traffic flow simulation.

Through demonstrations of xLSTM and diffusion models, Q.ANT shows that its hardware can run the most demanding categories of modern AI and is built for a broad range of AI use cases.

The ISC demonstration is the latest result in a series of third-party integrations, commercial collaborations, and institutional deployments for Q.ANT. In May, Q.ANT secured its first commercial orders through a partnership with German cloud service provider IONOS. In April, partner Daisytuner announced the development of a compiler using standard AI toolchains for real-time object detection applications. European high-performance computing facilities such as the Leibniz Supercomputing Centre in Munich and the Jülich Supercomputing Centre are running Q.ANT hardware in live production environments.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com