README.md

---
title: NuWave
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
license: agpl-3.0
---

# NuWave — Your Model Gets Smarter Over Time

CPU-native AI through compound context optimization.

Powered by **BitNet b1.58-2B-4T** — Microsoft's ternary weight model (1.58 bits per weight, addition-only inference, 0.4GB memory). No GPU needed.

## How It Works

- **KISS** (Keep Input Simple, Substrate) — filters redundant context. System prompt skipped when unchanged. Old conversation history compressed to summaries.
- **Pith** (The Living Context Lens) — manages the context window as a cache hierarchy. Strips clutter, evicts cold entries, promotes what's relevant.
- **BitNet** — ternary weights {-1, 0, +1} replace floating-point multiplication with addition. CPU-native by design.

Together they compound: cleaner input → less processing → better context → sharper reasoning → both improve.

## Demo

Two tabs:
1. **Live Chat** — talk to the model, see KISS/Pith metrics in real time (tokens saved, system skipped, clutter stripped)
2. **A/B Benchmark** — same conversation through baseline vs NuWave. Watch tokens drop and time decrease as KISS learns what's redundant.