Quentin Technologies is pioneering a fundamental shift in how language models learn and operate. We are developing Quentin, a byte-level language model that replaces traditional backpropagation with local, causal learning rules. Our model reads raw byte streams once, forward in time, predicting each byte before it arrives and dynamically growing or pruning its structure based on outcomes. This approach eliminates the need for tokenizers, separate training and inference phases, and frozen weight files—enabling models that keep learning after deployment.

Our mission is to make continual, local learning practical without requiring expensive datacenter GPUs or massive computational resources. We focus on compression efficiency, continual learning without catastrophic forgetting, and principled structural pruning under hard memory budgets. The result is a language model that operates within tens of megabytes rather than gigabytes, making it ideal for on-device assistants, edge deployments, and systems that must adapt after deployment. We're answering a fundamental research question: can capable language models be built around local learning and structural growth rather than large-scale gradient descent over static data?