Chinese AI Models Are Gaining an Advantage in Crypto Trading

Chinese artificial intelligence models DeepSeek and Qwen AI outperformed their American competitors in a cryptocurrency trading competition conducted with real money. This development was confirmed by the financial market-focused AI research laboratory Nof1.

Nof1 launched the Alpha Arena, a competition testing the investment capabilities of popular AI models, on October 17. The models were directed to trade cryptocurrencies on the decentralized exchange called Hyperliquid with a starting capital of $10,000, given the same questions and datasets initially.

During the competition, DeepSeek V3.1 Chat increased the initial $10,000 capital to $21,600, achieving a gain of %116. Qwen 3 Max, developed by Alibaba, also raised its capital by approximately %70, reaching $17,000.

When compared to previously tested models, Google's Gemini 2.5 Pro and OpenAI's ChatGPT 5 were found to have the lowest performance, with losses exceeding %60. Nof1 noted that these models generally preferred smaller positions, indicating they were less aggressive compared to their competitors.

Some experts suggested that the results of the Alpha Arena competition followed a random walk. This mathematical model describes a path consisting of consecutive random steps and posits that the average trading position tends to converge to the starting point.

The Alpha Arena competition will conclude on November 3, suggesting that significant changes in rankings may occur.

Alpha Arena is just one of many experiments testing the trading capabilities of AI models. In June, researchers from Stanford University found that a model trained only on public information outperformed %93 of managers over a 30-year period by an average of %600.

In July, a user initiated a six-month experiment selecting small stocks with ChatGPT and a starting capital of $100. Last week, the user's capital had decreased to $76. During the same period, investments in the S&P 500 would have increased to $109.46.