Stop Renting Intelligence: The Train-to-Deploy Loop for Specialized AI — Fireworks AI

This talk details the practical implementation of a continuous train-to-deploy loop for specialized AI using open-weight models. It highlights how customization, achieved through supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), and direct preference optimization (DPO), can bridge the quality gap with closed models. The presentation covers Fireworks AI's infrastructure, including the FireAttention engine with custom CUDA kernels for optimized inference, and speculative decoding with custom draft models. It outlines three training methods: a training agent, managed training, and a training API. A key focus is the 'flywheel' of production deployment, featuring one-click deploys, monitoring, and hot-loading of new weights with zero downtime, ensuring numerical fidelity between training and serving environments. Customer case studies (Cursor, Vercel, Genspark) are presented to demonstrate tangible benefits in inference speed, latency, and quality.