Concepts

Read these to understand why OpenTernary behaves as it does. The guides tell you how to do things.

• Ternary quantization

  ---

  What restricting weights to −1, 0 and +1 means, how scales work, per-tensor vs per-group, fake-quant vs packed, and why "1.58 bits" usually means 2.

• Runs and evidence

  ---

  Every result lives in a run directory. Why "completed" never means "good", how the quality gate works, and what the status symbols mean.

• Calibration

  ---

  How layer-by-layer reconstruction tries to win back quality: learned scales, learned thresholds and soft-to-hard assignments.

• Architecture

  ---

  How the CLI, services, adapters, backends, passes, exporters and evaluators fit together, and where plugins plug in.