Valory Unveils Open-Weight Olas Predict Model With Gains in Real-World Forecasting

What happened

Valory released Olas-Predict-Qwen3.8-27B, an open-weight AI model designed for forecasting in prediction markets. In internal tests, this 27-billion-parameter model achieved 67.3% balanced accuracy, nearly matching the 68.5% of Valory's GPT-5.6 comparator when both used the same evidence. The evaluation covered 6,824 prediction markets using a standardized prompt and evidence set. The model is available for organizations to deploy and adapt as needed.

Key details

• Olas-Predict-Qwen3.8-27B reached 67.3% balanced accuracy on Valory's test set.
• GPT-5.6 comparator scored 68.5% balanced accuracy under the same conditions.
• The evaluation used 6,824 prediction markets with a knowledge cutoff of February 16, 2026.
• The open-weight model has 27 billion parameters.
• Brier score for Olas-Predict-Qwen3.8-27B was 0.2124, compared to 0.2029 for GPT-5.6.

Why it matters

Organizations and developers can now use or modify a high-performing forecasting model without relying on proprietary services, giving them more control over deployment and updates. The open-weight release allows teams to tailor the model to their own infrastructure and evidence pipelines, but they will need suitable hardware to run it. The results also clarify the gains from specialized training versus starting with a stronger foundational model, which can inform future development and resource allocation for teams building forecasting systems.