OneBit AI - Company Profile

Developer of Extreme Ultra-low-bit Frontier AI models that run locally on the edge. OneBit AI builds ultra-low-bit AI models designed to run locally on CPUs and edge devices instead of relying on expensive cloud GPUs. Our core approach is to train models natively at extremely low bit-widths, including fully ternary models, rather than simply compressing conventional models after training. This allows us to explore a fundamentally different way of executing AI with dramatically smaller models and lower hardware requirements. We are building models and an inference stack across language, speech, and vision. Our current work includes a 70M-parameter fully ternary language model reaching ~600 tokens/sec on CPU, as well as 5–10 MB edge-native speech-to-text and text-to-speech models. We have also built an offline AI agent for document-audit workflows that operates without cloud connectivity. Our initial applications include coding agents, enterprise copilots, document processing, customer support, transcription, and other agentic workflows where privacy, latency, cost, or connectivity make cloud inference impractical. Long term, we believe this architecture can enable persistent AI at the operating-system level—local intelligence that can power applications and agents without sending every interaction to the cloud.