A 4B Open Model Matches GPT-5.6 Sol on Retrieval at 100× Lower Cost

What if one of the biggest problems with AI agents isn't that their models aren't powerful enough—but that we're using expensive general-purpose models for jobs that smaller models could learn to do?