A new fine-tuning system from Overmind transforms a company's agent production records into training data to build specialized open-weight models. A tuned 9B Qwen 3.5 model using this platform achieved 59.6% accuracy for exact contract clause quotes, surpassing GPT-5.6 Luna's 8.5% and the 4.2% score of an untuned Qwen model. In trials across 4,000 questions on more than 100 contracts, the tuned 9B model reached 91.6% F1 against the 82.0% score of the GPT-5.6 Luna flagship.

Overmind models achieve 7x better accuracy, 28x fewer hallucinations, and 20x cheaper usage than frontier models in biomedical, legal, and engineering tasks. Customers own the resulting model weights and can download, retrain, or roll them back. The platform is open source under the AGPL-3.0 license, and in biomedical extraction tasks, its small models outperformed frontier flagships 4x.