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

Jetashree Ravi, Tech Lead Manager for Applied ML at Fireworks AI, gave this 19-minute talk at the AI Engineer World's Fair 2026 in San Francisco. Her subject is how teams can move from closed models to open ones without losing quality, by training open models on their own data and serving them in a continuous loop.

After the intro, she sets out three truths at 1:27. Open models such as GLM 5.2 are ready for production, customization closes the quality gap that remains, and the AI stack is shifting toward training and serving as one continuous loop. At 2:22 she places Fireworks in that picture, and at 3:12 she gives its scale, which the description puts at about 30 trillion tokens and more than 10,000 companies. Customer results follow at 3:47. According to the description, Cursor's Composer 2 was trained and hosted on Fireworks with 13x faster inference, Vercel saw a 40x latency improvement, and Genspark got 12% better quality from reinforcement fine-tuning.

The inference section covers the FireAttention engine and its custom CUDA kernels (5:27), then speculative decoding with custom draft models (6:27). At 7:22 she lists three ways to train: a training agent, managed training (8:22) and a training API (9:27). The API handles SFT, RFT and DPO, which get their own chapter at 10:07.

From 11:07 she describes the data flywheel, which the description says includes one-click deploys and monitoring. Hot loading new weights without downtime comes at 12:42. At 13:22 she covers numerical fidelity, meaning training and serving produce matching numerics. The last stretch is demos: the training agent at 14:22, then managed training through to deployment at 16:17, before the wrap-up at 18:42.

The scale and customer figures come from Fireworks itself.