Bespoke Labs released Nimble, a model for typed decisions on text. It answers schema-based questions without generating reasoning. Each answer is returned with probabilities for all allowed choices. The model is trained on 2,676 curated examples. It runs on Mac with Apple Silicon or Linux with NVIDIA GPU. A merged model requires 18 GB for weights alone. It supports enums and booleans but not nested fields. Prompts are capped at 2,048 tokens. It cannot judge images or return free-form text. Performance depends on the training data domains. The repository includes code for serving and training. The base model is Qwen3.5-9B. A parallel scorer processes context once on Mac.

• 2,676 training examples
• 2,048 max tokens per prompt
• 90.1% accuracy on 324 held-out examples
• 18 GB unquantized model weights