OS-Software/Ternary-Bonsai-27B-unpacked-heretic-ja

This is a decensored version of prism-ml/Ternary-Bonsai-27B-unpacked, made using Heretic v1.4.0 with the Arbitrary-Rank Ablation (ARA) method using a LoRA adapter and row-norm preservation

Abliteration parameters

Parameter	Value
start_layer_index	10
end_layer_index	51
preserve_good_behavior_weight	0.6491
steer_bad_behavior_weight	0.0008
overcorrect_relative_weight	0.9874
neighbor_count	14

Performance

Metric	This model	Original model (prism-ml/Ternary-Bonsai-27B-unpacked)
Keywords	2/100	98/100
KL divergence	0.0243	0 (by definition)

Note: Performance testing, including the measurement of refusal rates, was conducted using Japanese datasets (harmless_alpaca_ja, harmful_behaviors_ja).

GGUF Version

GGUF quantizations available here

---

Ternary Bonsai 27B — Unpacked FP16 Safetensors

FP16 safetensors (HuggingFace format) of the Ternary Bonsai 27B model. This repo exists for users who want to run Ternary Bonsai with stock HuggingFace tooling or frameworks that don't yet support the packed ternary format. The 2-bit hybrid-attention kernels are currently in our forks of MLX, mlx-swift, and llama.cpp — once they land upstream, this unpacked version will no longer be needed.

We strongly recommend using the natively packed models instead. The packed format is where all the benefits of Bonsai come from — a 7.2 GB deployed footprint (down from 54 GB), 95% of FP16 intelligence retained, and interactive decoding on everyday laptops (26 tok/s on an M5 Pro). This unpacked FP16 version is full-size and does not provide any of those advantages.

For the optimized ternary release models (recommended):

• Ternary-Bonsai-27B-mlx-2bit — Ternary MLX for Apple Silicon
• Ternary GGUF (Q2_0_g128) for llama.cpp (CUDA, Metal, CPU)

For the phone-class variant:

• Bonsai-27B-mlx-1bit — 1-bit Bonsai 27B (~3.9 GB), fits an iPhone 17 Pro Max