Bridgewater's fine-tuned Qwen3-235B outperforms GPT, Claude

Bridgewater and Thinking Machines Lab fine-tuned an open-weight model, Qwen3-235B, and in the teams' internal evaluation the model reached 84.7 percent accuracy on six investor-oriented financial-document tasks while costing nearly 14 times less to run than the best frontier model they tested.

How did they get a model to understand investors' judgment?

They trained an open model on proprietary, expert-labeled examples and used a staged labeling loop to reduce expert time. The team built the fine-tune on top of Qwen3-235B using the Tinker platform, first collecting labels from outside contractors, then training a first model on those imperfect labels and having that model re-evaluate the same documents. Where the model and original label disagreed, Bridgewater investors corrected only those disputed cases, concentrating expensive expert effort on the highest-value examples.

The report says that training with those corrected examples produced the 84.7 percent accuracy figure, a step above the 78.2 percent accuracy the authors measured for the best tested frontier model.

How did the fine-tuned model compare to GPT, Claude and other frontier models?

In the team’s tests, basic prompts left frontier models at about 50 percent accuracy; carefully written expert instructions and a three-tier rating system raised those models into the mid-70s, but still below the authors’ 80 percent threshold for trustworthy deployment. The fine-tuned Qwen3-235B reached 84.7 percent in the internal evaluation versus 78.2 percent for the best frontier model the teams tested. The report also notes that GPT 5.4 costs 43 percent more than 5.2 while being only marginally more accurate, illustrating diminishing per-dollar improvements among the largest proprietary models.

The six tasks the teams used mirror routine investor judgments: for example, deciding whether a news article is relevant to a company executive, or whether a central bank document signals future rate direction. The teams implemented a three-tier label taxonomy: "relevant and interesting," "relevant but uninteresting," and "irrelevant." That taxonomy and the investor corrections were the secret sauce the authors credit for better performance.

Why it matters

The result shows proprietary corporate data and internal human expertise can still outstrip generic frontier models when turned into training signal. Bridgewater and Thinking Machines present a concrete example: a fine-tuned, open-weight model that the teams say hits 84.7 percent accuracy on investor-oriented tasks and costs nearly 14 times less to run than the best frontier model they tested. That implies firms with valuable private data can build domain-specialized systems without handing sensitive material to big labs, and that cost-performance trade-offs still favor targeted fine-tuning in some cases.

What are the methodological limits of the claim?

The numbers come from the collaborators' own internal evaluation. The report acknowledges this is not an independent comparison and both companies have a commercial interest in the result. The measurement choices matter: basic prompts for frontier models produced about 50 percent accuracy, while expert-written instructions raised their scores into the mid-70s but did not reach the team’s 80 percent deployment threshold.

What to watch

Look for independent replications or third-party benchmarks that apply the same six investor tasks and the three-tier label scheme. Also watch whether other firms deploy similar staged-labeling pipelines and whether open-weight fine-tunes on proprietary data close the gap in other high-sensitivity domains.