A new open source framework called Nimble provides a model and data curation recipe for creating decision-making agents similar to those from TypeSafe AI. The system is a LoRA finetune of a Qwen3.5-9B model that utilizes synthetic data across 10 categories to train the AI to discriminate between facts. It posted a 90% accuracy rate on a curated evaluation, compared to 66% for the original Qwen and 93% for Jev, and operates with 100ms latency on an H100 GPU.

The training methodology, termed contrastive data curation, generates negative data by slightly altering facts to improve model calibration without requiring probabilities in the training set. One industry observer claimed the release violated Jev's terms of use, though the Nimble weights were produced in two days of public development. Bespoke Labs noted that since no standard benchmark exists for this model class, Nimble's performance relative to Jev may vary on other metrics.