Kaiju

Character.ai’s in-house family of dense transformer LLMs (Small 13B, Medium 34B, Large 110B), built for fast, engaging, safety-aware conversation rather than raw academic benchmarks alone.

How it works

Kaiju models are dense (non-MoE) Transformers optimised for Character.ai's specific product requirements: fast inference for interactive character roleplay, high throughput for serving millions of concurrent users, and safety properties tailored to creative conversation. They use Character.ai's custom training infrastructure including the Squinch gradient compression algorithm and sticky session serving. Models are trained continuously on Character.ai's unique large-scale conversational data.

Why it matters

Kaiju represents a different model development philosophy: optimising for product fit and infrastructure efficiency rather than benchmark leaderboard performance. Character.ai serves more daily active users than many better-known AI products, demonstrating that inference-optimised models can achieve enormous commercial scale. Kaiju's architecture choices — dense rather than MoE, optimised for latency rather than raw capability — offer an important counterpoint to the frontier model race narrative.