Documentation

Welcome to the momo-kibidango documentation. Learn how to integrate and use 3-model speculative decoding to accelerate your LLM inference on Apple Silicon.

Key Features

• ✅ 1.97x faster inference with zero quality degradation
• ✅ OpenClaw native integration - works out of the box
• ✅ Memory efficient - runs on 16GB MacBooks
• ✅ Production ready - v1.0.0 with monitoring and metrics
• ✅ Smart fallback - gracefully handles edge cases
• ✅ MIT licensed - free for commercial use

How It Works

momo-kibidango implements Google Research's 3-model pyramid architecture:

1. 1. Draft Generation: Haiku 2 (fastest model) generates multiple draft tokens at 45.6 tok/s
2. 2. Middle Verification: Haiku 3 verifies drafts, correcting obvious errors at 30.5 tok/s
3. 3. Final Authority: Sonnet 3.5 validates the final output, ensuring quality matches baseline

This approach achieves near 2x speedup while mathematically guaranteeing identical output distribution to running Sonnet 3.5 alone.