GPU Internals, Turning Existing Corpora Into Training Tasks and Managing the RL Loop

A 4B open-source model post-trained with Castform matched GPT-5.6 Sol on retrieval accuracy while costing 100x less, according to Neon. The company says a typical multi-turn search request with gpt-5.6-sol takes over 10 seconds and costs about $0.03 end-to-end, making it slow and expensive. Castform enables developers to run reinforcement learning post-training without handling machine learning or GPU internals, turning existing corpora into training tasks and managing the RL loop.

During training, the agent repeatedly calls Lakebase Search, creating bursty workloads that Neon's dynamic compute scaling absorbs. Neon branching gives each rollout an isolated database state, and time-travel queries reconstruct agent states. Most teams' best training data is just sitting in their databases.

The problem is that turning raw data into something usable is hard, and letting agents read, search, and mutate data cheaply at scale requires advanced infra. Pointing Castform at Neon skips both.