Olas unveiled Olas-Predict-R1-14B on Tuesday after training it on 214,529 forecasts from 5,116 resolved prediction markets. The model achieved 75.8% accuracy across 2,628 previously unseen markets, compared with 75.4% for OpenAI's GPT-4.1 and 71.4% for the underlying DeepSeek model, according to benchmark materials. The models were not shown existing market odds during the test, so they could not simply copy the crowd. Fine-tuning improved the model's Brier score by roughly 20%, and that metric rewards accurate probabilities while penalizing confident mistakes. David Minarsch, CEO of Valory and a founding member of Olas, said the results indicate another path to AI performance besides larger general-purpose models. He said frontier models face increasing economic pressure from specialized systems and may need new markets to maintain their growth rates. Olas says the model can run on a single GPU and that its publicly available weights allow developers to run it themselves instead of paying a closed AI provider for each forecast. TrendForce expects the nine largest cloud providers to spend more than $886.7 billion in total capital expenditures in 2026, up roughly 90%, with AI infrastructure driving much of the increase. NVIDIA is central to that spending because larger models generally require more GPUs and data-center capacity. Minarsch stopped short of saying that spending is misplaced, noting that Olas-Predict itself was built on a model that required far more computing power to create. He said cheaper specialist systems could nevertheless pressure closed-model economics by handling particular jobs as well as larger systems. Thomson Reuters has also invested about $40 million to build a specialist model using an open foundation model and proprietary data, saying ownership provides greater control over performance, privacy and costs. In an experiment requested by the publisher, Olas-Predict gave NVIDIA a 70% chance of ending 2026 as the world's largest company by market capitalization, while Polymarket traders currently put NVIDIA at 78%. Olas says the model is available for download on Hugging Face, and its marketplace offers pay-per-call access.
Olas Model Matches GPT-4.1 Accuracy in Prediction Market Test
NS3Share
Summary
Olas-Predict-R1-14B hit 75.8% accuracy in price prediction tests, matching GPT-4.1’s 75.4% and outperforming DeepSeek’s 71.4%. Trained on 214,529 resolved markets, the model runs on a single GPU and offers open weights. CEO David Minarsch said specialized AI could disrupt big model economics. In one test, Olas-Predict gave NVIDIA a 70% chance to top market cap by 2026, below Polymarket’s 78%. The model also handles Bitcoin price prediction scenarios with high precision.
Source:Show original
Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information.
Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.