Uploaded November 2025 | Updated November 2026, 1 day ago
At Ray Summit 2025, Haoran Li from Character AI shares how the company powers its massive AI entertainment platform—serving tens of millions of users daily—using a fully open-source post-training stack built entirely on the Ray ecosystem.
Haoran next introduces Rayman, Character AI’s internal fine-tuning platform that accelerates model development velocity and significantly improves the training efficiency of large Mixture-of-Experts (MoE) models. Rayman brings together scalable distributed compute, efficient orchestration, and flexible experimentation—all built on open-source components.
Haoran also explains how Character AI leverages and adapts open-source RL libraries such as Verl to address the company’s unique reinforcement learning challenges, enabling rapid iteration and high-quality post-training for LLM-based interactive characters.
In this talk, she walks through the architecture of Rayman, the open-source projects that form the backbone of the system, the RL framework that drives model refinement, and the key ML challenges Character AI has overcome on its path to delivering state-of-the-art AI entertainment at global scale.
Attendees will gain practical insights into building scalable post-training pipelines, adapting RL for large consumer applications, and accelerating model improvement using Ray-based infrastructure.
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
At Ray Summit 2025, Haoran Li from Character AI shares how the company powers its massive AI entertainment platform—serving tens of millions of users daily—using a fully open-source post-training stack built entirely on the Ray ecosystem.
Haoran next introduces Rayman, Character AI’s internal fine-tuning platform that accelerates model development velocity and significantly improves the training efficiency of large Mixture-of-Experts (MoE) models. Rayman brings together scalable distributed compute, efficient orchestration, and flexible experimentation—all built on open-source components.
Haoran also explains how Character AI leverages and adapts open-source RL libraries such as Verl to address the company’s unique reinforcement learning challenges, enabling rapid iteration and high-quality post-training for LLM-based interactive characters.
In this talk, she walks through the architecture of Rayman, the open-source projects that form the backbone of the system, the RL framework that drives model refinement, and the key ML challenges Character AI has overcome on its path to delivering state-of-the-art AI entertainment at global scale.
Attendees will gain practical insights into building scalable post-training pipelines, adapting RL for large consumer applications, and accelerating model improvement using Ray-based infrastructure.
Liked this video? Check out other Ray Summit breakout session recordings youtube.com/playlist?list=PLzTswPQNepXllnU0C36WtkC0dqkAoDulh
Subscribe to our YouTube channel to stay up-to-date on the future of AI! youtube.com/c/anyscale
🔗 Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










