Neon and Castform Open-Source 4B Model for Cost-Effective Document Search

Neon and Castform have open-sourced a 4B parameter model trained via reinforcement learning that specializes in agentic document search and routing. The model matches or exceeds the accuracy of OpenAI's GPT-5.6 Sol in search tasks while reducing inference costs to approximately 1/100th. Agentic search is highly effective for complex queries but is typically slow and expensive when relying on frontier LLMs. This release demonstrates that small, specialized models trained with reinforcement learning can replace massive proprietary models for specific agent workflows, significantly lowering the barrier to deploying cost-effective AI agents. In Neon's evaluation, the 4B model achieved an average score of 1.447, outperforming GPT-5.6 Sol (1.369) and GPT-5.4 (1.377), with a single inference cost of $0.000929 compared to GPT-5.6 Sol's $0.087338. The training process evaluated the model on its ability to locate the correct document, cite the right passage, and deliver the correct final answer.