Bespoke Labs released the Nimble-9B model for response evaluation

Company Bespoke-Nimble-9B introduced a specialized LoRA adapter designed for automated evaluation of responses in boolean, enum, and rubric formats. The model is based on the Qwen3.5-9B architecture and optimized for classification and data structuring tasks.

Technically, the solution supports a context window of 8,192 tokens. To launch the adapter, you need to have the basic Qwen3.5-9B checkpoint and a graphics processor with CUDA support. In internal tests, the model demonstrated a result of 168 points, achieved in 2 hours of computation.

The tool is aimed at developers who need high accuracy when processing structured responses within agent systems. Using the LoRA adapter makes it possible to significantly reduce the requirements for computational resources compared to fine-tuning the full model weights, while maintaining specific classification skills.

The release is available in open access on the Hugging Face platform. Developers can integrate the solution into existing data processing pipelines where fast verification of LLM responses to given criteria or rubrics is required.