SprappSprappTernary (1.58-bit) language models compiled to WebAssembly. No server, no API key, nothing leaves your browser. Pick a model to download — your choice, on demand.
Choose a model to download
🏆 DISTILLED STORY — BEST QUALITY
Best story model. 40M ternary distilled from a 300M fp teacher (30B tokens). BPB 0.507 — beats TinyStories-1M (0.706) by wide margin. Fully offline.
📖 STORY MODELS
120M ternary model trained from scratch on TinyStories. Larger context, richer vocabulary. Fast, kid-friendly. Trained end-to-end, no distillation.
6.2M params, just 7 MB. Ternary distilled, 6k vocab. BPB 0.519. Best size-to-quality ratio. Instant load.
Every weight is a single bit {-1,+1}. Research demo: the floor of weight quantization, running in your browser.
💬 CHAT & INSTRUCTION
Strongest on-device chat — Qwen3-0.6B distilled to 1.58-bit ternary, runs fully offline. ChatML, instruction-following, reasoning. The flagship on-device assistant.
Full instruction-following chat — SmolLM2-360M distilled to 1.58-bit ternary. ChatML, follows complex prompts, summarizes, reasons. Lighter than Qwen3.
General chat with local search index (RAG, cites or abstains) and structured tool-calling. Grounded answers, works offline.
🗂 LEGACY
Superseded by meeny v6. BPB 0.685, same architecture, weaker teacher (1B, fewer tokens).
Downloads once, cached in your browser (IndexedDB) — works offline afterward. · clear cache