# Compact Bonsai 2 27B Philadelphia Class

**8.630 GB of weights plus correction:** original packed weights (8,595,477,990 bytes) + required rank-8 adapter (34,093,704 bytes). Tokenizer/config/runtime files are additional. Keep the two files separate; do not merge and requantize the adapter. The base file alone is unchanged original Bonsai and does not contain the abliteration.

The adapter approximates the previously validated mixed-Q8 derivative using randomized SVD of its weight differences from the original ternary model. It is a compressed approximation, not a bit-identical reconstruction or a newly trained model. All 126 edited output matrices retain rank-8 corrections in FP16. This avoids the failed method of rounding edited weights back to ternary.

## Validation

The actual original GGUF plus this adapter produced **0/126 refusals**, **100.0% usable openings**, and **21/24 benign capability passes**. Original Bonsai scored 123/126 refusals and 22/24 capability passes; the larger mixed-Q4/Q8 releases scored 0/126 and 21/24. These are limited heuristic screens, not general intelligence or safety guarantees. MLX passed strict model+adapter loading, three numerical layer checks and one arithmetic generation prompt on Linux CPU. See `evaluation.json` for the exact scope.

## Run

```bash
hf download KridgeDookie/Ternary-Bonsai-2-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS-MLX-Mixed-2-4bit --include "compact/*" --local-dir bonsai-compact
pip install -r bonsai-compact/compact/requirements.txt
python bonsai-compact/compact/compact_runtime.py --model bonsai-compact/compact --prompt "Say hello briefly."
```

The supplied loader is required: `from compact_runtime import load_compact; model, processor, config = load_compact(directory)`. It first loads the original schema-2 pack, then installs all 126 additive corrections. Loading just `model.safetensors` with the original loader omits the edits. This is the same compact 2-bit package in both mixed-precision MLX repositories; those repository names describe their older root-level models. The original vision weights are retained, but vision behavior was not tested. Linux CPU validation used the pinned MLX stack; macOS Metal has not been tested.

## Provenance

Created using Bonsai by Prism ML. Original base: [prism-ml/Ternary-Bonsai-2-27B-mlx-2bit](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit/tree/3f926b415992eaa2ae9dd7b573706494d6bbf787); mixed-Q8 reference revision `61231c9a2feed6457f47d1d50a691b8dc96b2613`. Apache-2.0 weights; source runtime retains its MIT license. The compact correction and loader were added 2026-09-21. `artifact_manifest.json` records hashes of the exact source and adapter files. `compression.json` records numerical approximation errors; `compress_adapter.py` records the build method.