Optimising Continuous RL at Fireworks: Rollout, Reward, Update, Repeat!

Learn how Fireworks.ai runs continuous reinforcement learning for LLMs in production at this online technical meetup on 17 August 2026 at 5:00 PM. The reinforcement learning loop sounds simple, rollout, reward, weight update, but keeping that loop effective in production depends on the underlying data plumbing and the design of the learning environment. This session from the AI Performance Engineering Meetup (London–Cambridge) examines the engineering decisions that determine whether a model actually learns during continuous post-training. ## Inside a production RL loop In the main talk, “Rollout, Reward, Update, Repeat: How Fireworks Optimises Continuous Post-Training,” Sinan Ozdemir of Fireworks.ai will walk through a post-training reinforcement learning session on Kimi K3 using Fireworks’ serverless API. The presentation will explore how key choices interact across the training process, including: - RL algorithms and loss functions - LoRA rank and hyperparameter selection - Reward design - Rollout, evaluation and weight-update stages - Data plumbing for continuous production workflows - The conditions that help a model learn, or prevent it from learning Rather than treating each setting in isolation, the talk focuses on how the complete learning environment shapes model behaviour. Attendees can expect a practical, systems-oriented view of continuous RL for large language models, grounded in a worked post-training session. ## Format and audience The event opens with introductions and meetup updates from Chris Fregly and Antje Barth, followed by Sinan Ozdemir’s technical presentation. It takes place online via Zoom as part of the group’s recurring third-Monday meetup series. This session is particularly relevant to AI and machine learning engineers, LLM practitioners, performance engineers, researchers, and technical builders interested in reinforcement learning, model post-training and production AI infrastructure. The subject matter is technical; no formal prerequisite level is specified on the event page.… Register for this event