Bespoke Labs Releases Nimble to Challenge 27B Models

Bespoke Labs has introduced Nimble, an open-weight recipe and repository designed for typed classification tasks. Built as an open alternative to TypeSafe's Jev model, the Bespoke-Nimble-9B model achieves a 90.12% agreement rate on a 324-example holdout benchmark, closely trailing Jev's 93.21% performance despite utilizing a much smaller parameter footprint. The model is trained using a Low-Rank Adaptation (LoRA) on the Qwen3.5-9B base model, requiring only a single epoch of training over 2,676 curated contrastive pairs.

To achieve high efficiency, Nimble bypasses traditional autoregressive text generation and chain-of-thought reasoning. Instead, it maps every allowed answer to a single vocabulary token and reads the model's logits directly. A softmax function converts these logits into a probability distribution, which Python then assembles into a typed response. This approach eliminates the risk of JSON parsing failures and removes the latency associated with generating explanatory text. The system supports flat schemas with no nested fields, where each field accepts either a Boolean or an enum of fixed strings.

For practitioners, Nimble offers local execution without API dependencies. It includes two distinct runtimes: an MLX-based ParallelScorer for Apple Silicon and a CUDA-based scorer for NVIDIA GPUs. On Apple hardware, the ParallelScorer processes the shared context once and scores fields in parallel, whereas the CUDA implementation processes each field independently. Both runtimes return the selected answer alongside raw candidate logits and normalized probabilities.

The training process relies on contrastive data curation rather than teacher probabilities to shape the model's decision boundaries. By pairing nearly identical examples where only one focus fact is flipped, the training forces the model to recognize which specific evidence dictates a decision. This methodology provides developers with a highly predictable, lightweight classification tool that runs entirely on local hardware.