A LoRA finetune of Qwen3.5-9B served as the foundation for Nimble, a new decision model released by Bespoke with its full data and training recipe. The model utilizes a synthetic contrastive data curation method that slightly alters facts to create negative samples, forcing the AI to better discriminate during decision making. This approach increased accuracy on a curated internal evaluation from 66% for the base Qwen model to 90%, while maintaining 100ms latency on H100 hardware.

The release provides a public alternative to Jev, which maintains a 93% accuracy rate on the same internal evaluation. Bespoke noted that the absence of a standard benchmark means performance could be lower on other tests. The model was developed over two days without the use of distillation or reinforcement learning and is available for free use on local devices such as Macbooks.