Bespoke Nimble 9B

Bespoke Nimble 9B is a LoRA adapter for Qwen3.5-9B, released by Bespoke Labs. It performs context-grounded classification: given a context and a schema, it assigns probabilities to the allowed answer options (enum choices, booleans, or integer rubric score levels). It scores the candidate answer tokens directly and does not generate reasoning or free-form text in the evaluated workflow.

The repository contains the adapter (about 165 MiB), tokenizer, prompt builder and reference inference code. The base Qwen3.5-9B checkpoint must be downloaded separately. For rubric fields it returns the selected integer and a probability-weighted expected score. A September 24, 2026 update raised the limits to 8,192 prompt tokens (longer prompts are rejected, not truncated) and up to 255 choices per field, and uses a softmax temperature of 1.0 without a separate calibration fit. The earlier checkpoint remains available as a revision, and a v2 model exists separately.

It is run locally and needs a CUDA GPU with BF16 support, using PyTorch and the pinned requirements. MLX and Linux loaders are described in the GitHub quickstart. The model card lists English as the language. A YouTube creator described it as a Jev-like open-source model that performs close to Jev, but that claim is not verified here.