c2 model

Character.AI trains its own conversational language models in-house rather than relying only on third-party APIs, and the c2 model is part of that lineage of proprietary systems built to power chat with its character personas. The company’s earlier public model updates were named C1, C1.1, and C1.2, each aimed at making replies feel more in-character, more coherent over long conversations, and more responsive, while its internal model family (known as Kaiju) spans multiple sizes so different features can trade off latency against quality. Character.AI has not published parameter counts, training data, or architecture details for its chat models, treating them as trade secrets the way most consumer AI companion products do.

The model is not available outside Character.AI’s own app and website: there is no public API, no downloadable weights, and no third-party hosting. It is tuned specifically for roleplay and companion-style conversation rather than general assistant tasks like coding or document analysis, and Character.AI’s product decisions, such as response length, filtering, and memory features, matter as much to the user experience as the underlying model itself.