Bespoke Labs AI launched Bespoke Nimble, an open-source decision model designed to replicate Jev's capabilities using the Qwen3.5-9B architecture. The team developed the model in two days via LoRA finetuning on 2,676 training samples, reaching 90.12% accuracy on a 324-sample internal test set compared to 93.21% for Jev. The project avoided distillation from Jev and utilized contrastive data curation, which involves modifying specific facts to flip correct answers to teach the model how to isolate critical evidence.

Nimble optimizes for speed by reading candidate token logits to create structured output, bypassing the generation of Chain-of-Thought or JSON. This approach achieves 100ms latency on H100 GPUs. The release follows a three-year development period for the proprietary Jev model by an OpenAI co-founder, demonstrating the current speed at which open-weight models can replicate specialized decision-making architectures.