Olas Released Forecasting Model With 75.8% Accuracy
The 14-billion parameter model improves on its base architecture by providing more accurate probability estimates.
Updated on Sept. 22, 2026 in Artificial Intelligence
Live Poll
Do you trust specialized, smaller AI models to perform as reliably as large general-purpose systems?
Olas has released Olas-Predict-R1-14B, an open-weight forecasting model that achieved 75.8% accuracy on prediction-market tests. The model is a fine-tuned version of the DeepSeek-R1-Distill-Qwen-14B architecture designed to operate on a single GPU.
Why it matters
This release provides a specialized alternative to general-purpose LLMs by leveraging a verified dataset of real-world outcomes. The model aims to bridge the performance gap in automated forecasting, which is critical for decision-making in volatile environments.
The 14-billion parameter model achieved a 75.8% accuracy rate on 2,628 test markets, representing a 20% relative reduction in forecasting error compared to the base DeepSeek-R1-Distill-Qwen-14B model.
The players
Olas
An AI development entity that maintains the Olas Predict forecasting platform and focuses on open-weight model architectures.
The details
Olas-Predict-R1-14B is trained on 214,529 prediction records derived from 5,116 resolved markets. By fine-tuning the base model on these verified real-world outcomes, the developers enable the system to ingest external data points and output refined probability estimates. The architecture is optimized for inference on a single GPU, contrasting with larger, more generalized models that often require multi-node clusters.
Timeline
2023: Olas Predict began operating.
September 22, 2026: The Olas-Predict-R1-14B model was released.
The Tech Race
This release follows the industry trend of fine-tuning open-weight models for narrow, high-utility domains like probabilistic forecasting. It seeks to outperform general-purpose base models by prioritizing historical accuracy on resolved real-world events.
Developers can now run a specialized forecasting model on standard single-GPU hardware configurations. This provides a more accessible path for integrating predictive market logic into automated workflows compared to larger, proprietary AI systems.
The takeaway
Olas has demonstrated that domain-specific fine-tuning can significantly reduce forecasting error in smaller parameter models. Observers should track the release of updated models to see if this accuracy gain holds as the volume of training market data scales over time.
What happens next
Olas plans to continue training new iterations of the forecasting model as more market records become available for inclusion in the training dataset.
Further reading
For more developments in specialized model architectures, visit Artificial Intelligence.
Source note: This article includes information reported by IT Brief Australia.
Live Poll
Do you trust specialized, smaller AI models to perform as reliably as large general-purpose systems?
