Castform and Neon Achieve GPT-5.6 Sol-Level Retrieval Accuracy with 100x Cost Reduction Using Open-Source Models

Castform and Neon developed an open-source 4B model that matches GPT-5.6 Sol's retrieval accuracy while reducing costs by 100x through reinforcement learning post-training. The solution leverages Neon's Lakebase for efficient data retrieval and infrastructure scaling.

Comments

Read original article