frontier models failed finance tests until open weights + private data
a fine-tuned qwen3-235b outperformed gpt and claude on bridgewater's financial analysis tasks at 14x lower cost, proving proprietary data is the real moat.
bridgewater and thinking machines lab fine-tuned the open-weight qwen3-235b model on internal financial data, achieving 84.7% accuracy on document triage tasks compared to 78.2% for top proprietary models. the project highlighted that frontier models like gpt and claude failed to meet the 80% deployment threshold because the necessary 'judgment' data exists only in private corporate archives, not public training sets. the fine-tuned model operates at nearly 14 times lower cost than commercial alternatives, offering a blueprint for companies to build competitive ai without sharing sensitive trade secrets with large labs.