Australian-listed chip designer BrainChip has launched Sequora, a processor design for running generative AI, including large language models (LLMs), directly on battery-powered devices such as earbuds and sensors.
BrainChip announced Sequora on 14 October 2026. It does not sell Sequora as a finished chip. Instead it licenses the design, called intellectual property or IP, which other companies build into their own chips. BrainChip says it is “available now as licensable semiconductor IP”.
Why the memory stays fixed
Most of today’s language models are transformers. As a conversation gets longer, a transformer stores more and more information about earlier words in a memory area called the KV cache (key-value cache). That memory keeps growing, and so does the work needed for each new word. This is a problem on small devices with little memory and a small battery.
Sequora uses a different kind of model called a state-space model. Rather than storing every earlier word, it carries a fixed-size summary, or “state”, that it updates as each new piece of data arrives. BrainChip says this design “keeps memory use fixed no matter how long the data stream or conversation runs”.
BrainChip says Sequora can run models from a few million parameters up to 3 billion.
The claims
The efficiency claims are BrainChip’s own. The company did not publish benchmarks, speed figures or power measurements.
BrainChip says Sequora can work alongside Akida, its existing low-power chip design for processing sensor data such as audio, vibration and radar. “Customers are not asking for just a signal classifier or a language model,” said Sean Hehir, BrainChip’s chief executive. “They want both, in the same product, on the same battery.”
Frank Ferrante, chief executive of chip company ForwardEdge ASIC, said in the release that Sequora “gives us a path to add on-device intelligence to mission-critical edge ASICs”.
What this means
- AI is moving onto small devices. Running models locally can keep data on the device.
- Fixed memory is the selling point. State-space models avoid the growing memory of transformers.
- No numbers yet. Without benchmarks, BrainChip’s efficiency claims cannot be compared.
Sources
- BrainChip, “BrainChip Launches Sequora: Edge AI That Doesn’t Just Recognize — It Understands and Responds”, GlobeNewswire, 14 October 2026. https://uk.advfn.com/stock-market/TG/24Y/share-news/BrainChip-Launches-Sequora-Edge-AI-That-Doesnt-Ju/99490278
