GitHub Trending · AI Apps & Models · Updated
PrismML-Eng/Bonsai-demo
Local-run demo of Bonsai, the 1-bit / ternary ultra-compressed LLMs
- Stars
- ★ 1.8k
- Language
- Shell
- Chart appearances
- 1
- On-chart tracking
- 2 days tracked
What it does
The official demo repo for PrismML's Bonsai model family, built around extreme quantization: the 1-bit version uses ~1.125 bits per weight so a 27B model fits a modern iPhone; the ternary version (~1.7 bits) is the higher-quality default. The newest 27B is a vision-language model with image input, tool calling, MCP, 256k+ context and adjustable reasoning effort.
What's inside
A pure Shell project: setup.sh / setup.ps1 one-liners install dependencies and pull models, scripts/ holds run/download scripts, plus three whitepaper PDFs, an AGENTS.md (a config guide written for AI agents), topical docs (VISION/TOOLS/…) and community benchmarks.
Tech stack
Shell + a llama.cpp server, models hosted on HuggingFace; runs on Mac Metal, CUDA/Vulkan/ROCm or plain CPU; Apache-2.0.
Use cases
- Mobile / edge developers — run a 27B vision-language model offline on an iPhone
- Privacy-sensitive settings — local AI assistants for medical or legal data that can't leave the device
- Hardware tinkerers — run LLMs on low-power boards like a Raspberry Pi
Why it's trending
The vision-language model and on-device iPhone demo make a technically ambitious claim concrete, while the included scripts reduce trial friction.
Current chart placements
| Period | Language filter | Rank | Stars gained |
|---|---|---|---|
| Daily | All languages | #7 | +278 |
Editorial summaries are based on the public repository and the dated GitHub Trending snapshot. Verify implementation details in the upstream repository.