Show HN: Gutsy – small CPU model yes/no/choice questions, get probabilities back (huggingface.co)

Gutsy, a new AI model, has been released that allows users to input state information and receive calibrated probability assessments for yes/no and choice-style questions, all executed locally on standard CPUs. The model, a fine-tuning of Qwen3.5, is available as both a 775 MB Q8_0 GGUF and a 505 MB Q4_K_M GGUF that operates roughly 25-30% faster while maintaining accuracy. This local, deterministic solution avoids the costs and privacy risks associated with networked AI models, making it a significant advance for applications requiring private and efficient decision-making. Key technical features of Gutsy include its robust option-order handling, state caching for efficiency in follow-up questions, and the ability to process up to 255 options, providing a reliable metric for "none of these" outcomes when applicable. Calibrated using proper scoring rules, the model boasts a validation calibration error of 0.022, indicating a strong alignment between predicted probabilities and actual outcomes. Overall, Gutsy marks a promising step towards more practical and controlled AI decision-making applications, particularly in contexts where privacy and computational efficiency are paramount.