Show HN: Tiny model for fast typed decisions on CPU (Gutsy)

Category: ai-ml

Tags: decision-making, calibration, local-inference, cpu, llama-cpp, fine-tuning

Score: 7.5/10 (Innovation: 7, Technical: 8, Documentation: 8, Utility: 7)

Gutsy is a compact 0.8B-parameter model fine-tuned for local, calibrated decision-making from text or JSON states, returning probabilities for yes/no, choice, or score questions. It runs on CPU via llama.cpp and offers a Jev-compatible API, making it practical for privacy-sensitive, cost-free inference. The project stands out with strong calibration, deterministic outputs, and extensive benchmarks, though it has known weaknesses in arithmetic and complex reasoning.

Target audience: backend devs, data engineers, ML engineers, product teams needing local decision models

Repository: https://github.com/kouhxp/gutsy · HTML · Apache-2.0 · 94 stars

View on Hacker News