bonsai image ternary 4b mlx 2bit.stub

---

id: bonsai-image-ternary-4b-mlx-2bit name: Bonsai Image Ternary 4B MLX 2-bit repo_path: ~/Documents/sw30labs/repos/bonsai-image-ternary-4b-mlx-2bit remote_url: https://github.com/sw30labs/bonsai-image-ternary-4b-mlx-2bit.git primary_language: Python framework: MLX (Apple Silicon) category: local-inference-mlx stacks: [mlx, cli] ingested: 2026-05-29 last_commit_date: 2026-05-28

Purpose

Local Apple Silicon wrapper for prism-ml/bonsai-image-ternary-4B-mlx-2bit, a ternary-quantized 4B image generation model running via the PrismML runtime. Adds a CLI and zero-build browser GUI on top of the upstream PrismML-Eng/Bonsai-Image-Demo runtime.

Architecture

• Runtime: Vendors PrismML-Eng/Bonsai-Image-Demo under runtime/; requires full Xcode Metal toolchain (not just CLT)
• CLI: Small Python CLI in bonsai/ for headless inference
• GUI: Zero-build browser GUI (gui/, gui.sh) for interactive image generation
• Install: install.sh -y handles env setup and model download
• Model format: 2-bit ternary quantization for reduced Apple Silicon memory footprint

Dependencies

• MLX / PrismML runtime
• Xcode + Metal toolchain (xcrun metal required)
• pyproject.toml-based packaging

CI/CD

None.

Entities and topics

• local-inference-mlx
• Tags: [mlx, apple-silicon, local-inference]