Olas has released a new specialized AI model, Olas-Predict-R1-14B, which matches the performance of GPT-4.1 in forecasting real-world events while operating on significantly less computing power.
Key Points
- Olas-Predict-R1-14B achieved 75.8% accuracy across 2,628 prediction markets, outperforming the 75.4% accuracy of GPT-4.1.
- The model was trained on 214,529 forecasts from 5,116 resolved markets to improve predictive probability and reduce confident errors.
- The system is designed to run on a single GPU, offering a cost-effective alternative to large-scale, closed-source general-purpose AI models.
- Weights for the new model are publicly available on Hugging Face, allowing developers to host the technology independently.
- Thomson Reuters is pursuing a similar strategy by investing $40 million into proprietary, specialized AI models to control costs and data privacy.