Thinking Machines partners with Bridgewater to build AI model that cuts errors by nearly 30%

When the world’s largest hedge fund decides its analysts are spending too much time on document busywork, it doesn’t just buy a ChatGPT subscription. Bridgewater Associates teamed up with Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, to build a custom fine-tuned model that reduces errors by 29.8% compared to the best available frontier models.

The results, published June 30 by Bridgewater’s AIA Labs and Thinking Machines Lab, show the specialized model hitting 84.7% average accuracy across six information-filtering tasks. Leading models like GPT, Claude, and Gemini variants, even when juiced with expert prompt engineering, were stuck in the mid-70s.

How they built it

The model was constructed on the Qwen3-235B base and trained using Thinking Machines’ proprietary Tinker platform. Two training techniques did most of the heavy lifting: interleaved batching delivered a 12.1% accuracy boost, while on-policy distillation added another 3.1%.

The Tinker API handled the infrastructure side, letting the team iterate rapidly without managing GPU clusters directly.