XNOR develops a proprietary method to render machine learning models as operations that can run quickly on nearly any processor, enabling real-time tasks on extremely limited hardware. Its technology delivers large speed, memory and power savings so devices with bargain-bin CPUs (even Raspberry Pi Zero–class hardware) can perform object recognition, tracking and voice recognition without cloud connectivity. Running models on-device reduces latency and preserves user privacy by avoiding sending data to third-party servers. The company offers off-the-shelf pre-trained models for common use cases and platforms as its Phase 1 product, with a planned launch in the fall. Phase 2, targeted for early 2019, will add on-demand customization for proprietary hardware or specialized recognition needs. Phase 3 aims to adapt cloud models for edge deployment, though no timeline was provided. Xnor.ai emerged as an official spin-off from the Allen Institute for AI to commercialize research that makes machine-learning models far less computationally intensive. Its core technology uses mathematical techniques to reduce the compute load for tasks such as object and speech recognition by an order of magnitude or two, enabling deployment on small devices. The underlying research is open source, though the company says the commercial product is not yet plug-and-play. Leadership includes CEO Ali Farhadi and CTO Mohammad Rastegari, and Madrona’s Matt McIlwain will join the board alongside AI2 CEO Oren Etzioni. Xnor.ai has begun commercialization efforts and has secured seed funding to advance those plans. Financially, the company raised $2.6 million in seed funding from its parent (AI2) and Madrona Venture Group.

Total raised

$15M

Funding rounds

2

Latest round

Series A

Latest activity

May 2018