.svg?fit=max&auto=format&n=mQScX1J5vEe0-L40&q=85&s=24c5c941d6f990a4aab4041c9b243291)
Technical references
Bonsai collection
All Bonsai models repositories: 1-bit and ternary, GGUF and MLX.
1-bit and Ternary Bonsai 27B whitepaper
Architecture, training approach, and evaluation methodology for the 27B family.
1-bit Bonsai 8B whitepaper
Architecture, training approach, and evaluation methodology for the 1-bit 8B.
Ternary-Bonsai 8B whitepaper
The 1.58-bit representation, group-wise scaling, and benchmark results.
Source code
| Repository | What it is |
|---|---|
| PrismML-Eng/Bonsai-demo | Setup scripts, run/server wrappers, and everything the Quickstart uses |
TOOLS.md (in the demo repo) | Tool calling and MCP server setup for Bonsai 27B |
AGENTS.md (in the demo repo) | Hardware-tuning knobs, written for AI coding agents helping someone set up the demo |
| PrismML-Eng/llama.cpp | llama.cpp fork (prism branch) with ternary Q2_0 kernels and pre-built binaries |
| PrismML-Eng/mlx | MLX fork with 1-bit support, until mlx#3161 merges upstream |
| huggingface.co/prism-ml | All model weights |
Community benchmarks
Measured throughput on real hardware (RTX 3080, GB10, Strix Halo, M4 Pro, and more) is collected incommunity-benchmarks/ in the demo repo. To contribute numbers for your hardware, copy the llama.cpp or MLX template in that folder, run the documented procedure, and open a pull request.
Get help
- Discord: discord.gg/prismml for questions, hardware reports, and announcements.
- GitHub issues: Bonsai-demo issues for bugs in the scripts or binaries.
- X / LinkedIn: @PrismML and PrismML on LinkedIn for release news.