Q.ANT Puts Light to Work Inside a Language Model
An optical processor and AMD GPU jointly run a roughly billion-parameter model, opening a practical test of a different approach to AI computation.
Published November 18, 2026Intelligence / Frontier Tech / Experimental Systems

A language model’s electricity bill begins with its mathematics. Every answer requires calculations, and changing how those calculations are performed could give developers another way to make AI affordable. German photonic-computing company Q.ANT is exploring that possibility with a processor that uses light to perform part of the work.
At the SC26 computing conference in Chicago on November 17, Q.ANT announced a hybrid language-model demonstration pairing its new Gen 2+ optical processor with an AMD GPU. The optical hardware handles Fourier transforms, mathematical operations that the company is using to explore alternatives to computationally intensive attention layers. Q.ANT announcement
Q.ANT’s public whitepaper description supplies the scale: a model with approximately one billion parameters, running inference jointly on an AMD Instinct MI210 and the photonic accelerator. Inference means using a trained model to generate results. Parameters are the numerical values a model learns during training. Whitepaper catalog
For developers, the intriguing possibility is a change in design freedom. Software could be organized around operations that optical hardware performs naturally. Whether that produces a better AI service will depend on the quality, speed and energy consumption of the complete system.
A different way to combine information
Attention helps language models account for relationships between pieces of text. Fourier-based approaches offer another way of mixing information across a sequence.
The mathematical idea has precedents. Google researchers’ FNet work, published at NAACL in 2022, replaced self-attention in a Transformer encoder with Fourier transforms. It achieved 92–97 percent of comparable BERT models’ accuracy on the GLUE language-understanding benchmark. That research established a useful algorithmic tradeoff on electronic hardware; its results are separate from Q.ANT’s new optical demonstration. FNet paper
Q.ANT’s contribution is an optical implementation of the transform inside a hybrid system. The company says Gen 2+ performs the operation through the physical behavior of light and retains PCIe connectivity for conventional servers.
“Photonic computing gives software developers a chance to rethink the mathematics behind their applications,” said Utz Bacher, Q.ANT’s vice president of software. Q.ANT announcement
From demonstration to useful workloads
A separate November 9 pilot with hosting company IONOS gives selected developers access to Q.ANT’s second-generation hardware in Germany. Participants can evaluate their own applications, with technical support and a final review of results. That program provides a route to testing useful workloads, although it should not be confused with commercial availability of the new Gen 2+ system. IONOS pilot
Q.ANT plans commercial systems combining Gen 2 and Gen 2+ processors for the second half of 2027. Q.ANT announcement
The public materials reviewed here do not establish an end-to-end speed or energy advantage for the language-model demonstration. The detailed whitepaper requires a request and was not examined.
The next useful evidence would compare equivalent tasks at a stated output quality, counting the GPU, optical hardware and supporting system together. A successful result could give AI developers another kind of processor to work with, and another reason to reconsider the mathematics their software asks computers to perform.