Q.ANT and AMD Run a 1B-Parameter LLM With an Optical Chip in Place of Attention
The longer the text in an LLM prompt, the faster compute and GPU memory for the KV-cache (the model's memory of what it has already read) grow. Q.ANT and AMD replaced some of the attention layers in a 1B-parameter model with an optical chip that computes a Fourier transform using light. Speed and quality figures are not disclosed, and they were not the first: British startup Lumai ran an LLM on optics in April. For developers this is still a demo, and it is too early to change your stack.
AI-processed from Habr AI; edited by Hamidun News
Picture a chatbot that answers questions over hundreds of pages of contracts, or an agent with a long conversation history. The longer the text a large language model (LLM) receives, the harder its job gets. The attention mechanism, in which every word is checked against every other word, needs quadratically more computation.
The KV-cache, the area of GPU memory where the model stores what it has already read, keeps growing and crowds out everything else. German company Q.ANT has tried to get around this limit with light.
At the SC26 conference in Chicago, it showed, together with AMD, hybrid inference (running a finished model) on a model of roughly 1B parameters. Part of the attention work there is done not by a GPU but by a photonic coprocessor. If you pay for GPUs to handle long contexts, this is worth knowing about, but there is nothing to buy yet: there are no speed or quality figures.
What exactly was shown
Some of the model's layers were moved from the GPU to the Q.ANT NPU Gen 2+ photonic coprocessor. It computes a discrete Fourier transform (a mathematical operation that mixes information between words by a fixed formula) optically, in a single pass of a light signal. According to the technical report, the experiment ran a language model of roughly 1B parameters on an AMD Instinct MI210 accelerator paired with the photonic chip. Utz Bacher, Q.ANT's vice president of software, calls the approach "Fourier attention layers" and describes it as an early prototype.
In everyday terms, the idea works like this. Ordinary attention is like a meeting where every participant talks to every other participant in turn. A Fourier transform mixes information according to a ready-made scheme, with no pairwise comparisons. That makes it cheaper, and on optics it is computed almost instantly. All the other transformer layers stay on the conventional AMD GPU.
What the company did not disclose, and why "world's first" is disputable
The press release calls the launch "the world's first," but back in April 2026 British startup Lumai ran a full LLM on its Iris server. Its architecture is different: instead of flat silicon waveguides like Q.ANT's, it uses three-dimensional optics with lenses and beams of light in free space. According to Lumai, models at the level of Llama 70B run in real time with 90% energy savings.
Q.ANT's demonstration has more gaps than data:
- The specific model is not named. It is unclear how the causal mask problem (forbidding the model from "peeking" at future words during step-by-step generation) was solved, or which layers remained on the GPU.
- There is no perplexity (a measure of how "surprised" the model is by a text) and no results on standard benchmarks. We are waiting for tests on Artificial Analysis.
- Neither the gain in tokens per second nor the actual KV-cache compression is disclosed. The reduction in energy use is described only as "potentially significant."
Earlier, in June 2026 at the ISC conference in Hamburg, Q.ANT showed a 59M-parameter diffusion model for image generation and the TiRex architecture based on xLSTM, running on the previous-generation chip. The company then claimed 30x energy efficiency, but that referred to the optical circuit itself, and the figures were calculated estimates.
What will happen to quality
Replacing attention with a Fourier layer was proposed even before photonics: in 2021 Google did it in the FNet architecture. Training sped up by 70–80%, but the model retained only 92–97% of BERT's accuracy on the GLUE benchmark. Moreover, FNet was an encoder, not a generative model that writes text word by word.
The optical version has the same architectural limitations plus overhead. At every layer, tensors must be converted from digital to light and back, and also shuttled between the GPU and the coprocessor. If those transfers eat up the gain, there will be no speedup.
What developers should do
Q.ANT plans to bring production systems to market in the second half of 2027. Previous-generation accelerators have been available since November 9 in a closed beta on dedicated IONOS servers in Germany. The chips have also been shipped to the German supercomputing centers LRZ and JSC, but no independent public reports from them exist yet.
You can experiment with the idea right now: in September 2026 the company released an open SDK with a C and Python API and CPU emulation. It is being updated for the Gen 2+ Fourier layers.
Two numbers will show whether the hybrid is viable: perplexity compared with an identical 1B-parameter model using classical attention, and an honest tokens-per-second gain on real inference. Without them, it is too early to talk about replacing attention. The open question is which will prove harder: bringing optical hardware to mass production or persuading developers to rewrite their models for it.
- Nov 19, 2026The Word "Video Game" Raised GPT-6 Astra's Willingness to Pick Targets From 6% to 92%
- Nov 19, 2026AWS adds a leaderboard of 70+ models to SageMaker: quality, speed and price in one table
- Nov 19, 2026Rubin delivers 50 PFLOPS per chip, TPU 8t 12.6 — but Google wins on scale
- Nov 18, 2026Huang steps down as NVIDIA CEO, company loses more than $1 trillion in market cap
Key Questions Answered
What exactly was shown?
Some of the model's layers were moved from the GPU to the Q.ANT NPU Gen 2+ photonic coprocessor. It computes a discrete Fourier transform (a mathematical operation that mixes information between words by a fixed formula) optically, in a single pass of a light signal. According to the technical report, the experiment ran a language model of roughly 1B parameters on an AMD Instinct MI210 accelerator paired with the photonic chip.
What the company did not disclose, and why "world's first" is disputable?
The press release calls the launch "the world's first," but back in April 2026 British startup Lumai ran a full LLM on its Iris server. Its architecture is different: instead of flat silicon waveguides like Q.ANT's, it uses three-dimensional optics with lenses and beams of light in free space. According to Lumai, models at the level of Llama 70B run in real time with 90% energy savings. Q.ANT's demonstration has more gaps than data: The specific model is not named.
What will happen to quality?
Replacing attention with a Fourier layer was proposed even before photonics: in 2021 Google did it in the FNet architecture. Training sped up by 70–80%, but the model retained only 92–97% of BERT's accuracy on the GLUE benchmark. Moreover, FNet was an encoder, not a generative model that writes text word by word. The optical version has the same architectural limitations plus overhead. At every layer, tensors must be converted from digital to light and back, and also shuttled between the GPU and the coprocessor.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.