Fireworks meetup: how to tune continuous RL post-training on Kimi K3

anbarth · x · 2026-08-17

The AI Performance Engineering Meetup hosts Fireworks.ai's Sinan Ozdemir for a talk on continuous post-training RL on Kimi K3 via Fireworks' serverless API.

The session covers algorithm choices, loss functions, LoRA rank, hyperparameters, and reward design—and how they interact with whether a model actually learns. The key to running RL continuously in production, per the abstract, lies in data plumbing and learning-environment design.

Original post →

More from Companies & People

Companies & People channel →