Bridgewater Associates’ AI application lab (Bridgewater AIA Labs) teamed up with Thinking Machines to fine-tune a relatively small model for financial information screening — and it’s outperforming the biggest frontier models on the market, including GPT-5.5 and Claude Opus 4.8. Across six everyday investment tasks like financial article relevance and central bank document analysis, the fine-tuned model cut error rates by nearly 30%, and its inference cost is roughly one-fourteenth that of the frontier models — a 13.8x reduction. The research highlights how domain-specific fine-tuning can give companies a "differentiated intelligence" edge.
The team used the Tinker platform to fine-tune the Qwen3-235B base model. Because labels from non-specialist annotators were riddled with errors, they built a verification and correction workflow: hard cases where the model’s prediction disagreed with the label were sent to investment experts for review. The final model’s accuracy jumped from a baseline of 44.8% to 84.66%, beating GPT-5.5’s 78.2% and Claude Opus 4.8’s 78.0%. And along with that significant accuracy boost, inference costs saw a staggering 13.8x drop.
Three core improvements drove the fine-tuning: first, interleaved batch processing, which alternates training across tasks in sequence to avoid interference from mixed training; second, a loss function with asymmetric clipping to optimize sampling; and third, online policy distillation, which penalizes the model when it drifts from the teacher distribution and dynamically updates the teacher model whenever validation performance hits a new high (every 20 steps).