license: apache-2.0
base_model:

• Qwen/Qwen3.8-27B
• prism-ml/Ternary-Bonsai-2-27B-gguf
• OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF
  tags:
• ninfer
• ternary
• bonsai
• heretic
• uncensored
• mtp
  pipeline_tag: text-generation
  language:
• en
• zh

Bonsai 2 27B — Uncensored, T5 NInfer pack

Model card for a local NInfer artifact:

Bonsai2-T5-MTP-r3-Q4-Uncensored.ninfer

Size 6,182,360,576 bytes. Written 2026-10-30 15:15. The repack report for that file is marked verified. The weight file in this repository is Bonsai2-T5-MTP-r3-Q4-Uncensored.ninfer.

SHA-256: 75125dacee2b6ccca8a2a6f318885ae9deb107e4e9e3dac3cd567e92b467cedb

This is not emiltsoi/Ternary-Bonsai-2-27B-Uncensored-Heretic-NInfer. That card describes a different artifact of about 9.52 GB with a vision tower and a DFlash2 adapter. The numbers and launch example on that card do not apply to this file.

Important notice

The language weights come from OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF. That release states that safety alignment was substantially reduced. The model is more likely than the official Bonsai weights to produce harmful, inaccurate, biased, offensive, or otherwise inappropriate content.

OS-Software describes the intended use as research and experimentation, including safety research, alignment studies, and red-teaming, and asks that it not be deployed as a public end-user service. Treat outputs as untrusted and check them before use. The publisher of those weights provides them without warranty and disclaims liability for misuse. Use is at your own risk, and you are responsible for complying with applicable law and the upstream licenses.

The refusal and perplexity figures later in this card were published by OS-Software for their GGUF release. They were not remeasured on this repacked NInfer file.

What this file contains

The finished artifact has two components: the text model and one MTP head. The conversion record describes a Qwen3.5-style text stack: 64 layers, hidden size 5120, vocabulary 248,320, and a position limit of 262,144. About three quarters of the layers are gated linear attention and about one quarter are full attention.

The intermediate PTQ1 NInfer file contains 790 objects: 322 ternary groups (ptq1_g128_fp16), 5 Q4_0 MTP matrices, 360 BF16 tensors, and 96 FP32 tensors. The T5 repack keeps those bindings and changes how the ternary weights are stored.

This file does not contain a vision tower or a second draft adapter. Image input needs a separate vision projector.

Provenance

1. The Qwen team released Qwen/Qwen3.8-27B under Apache-2.0. The local qwen3.8-27b tree supplied the tokenizer, chat template, and configuration.
2. Prism ML compressed the language weights to ternary values without changing the architecture. Each group of 128 weights has an FP16 scale, and a block Hadamard rotation is folded into the stored weights. The official PTQ1_0 packing is about 1.75 bits per weight. The model card lists a context limit of 262,144. License: Apache-2.0. Card: prism-ml/Ternary-Bonsai-2-27B-gguf. Notes: Ternary Bonsai 2 27B.
3. OS-Software baked a Heretic low-rank edit into the official ternary codes and left the original scales unchanged. Their card says 34 matrices changed, the remaining tensors were untouched, and the merge is approximate rather than an exact floating-point LoRA merge. The language weights in this pack come from their PTQ1_0 GGUF. That repository also declares Apache-2.0.
4. The MTP head was not edited with the language weights. The conversion script reads the existing r3 Q4 head from Ternary-Bonsai-2-27B-PTQ1_0-MTP-r3-Q4_0.gguf. model_mtp.safetensors, dated 2026-09-25, is also listed in the conversion sources.
5. On 2026-10-30 the pack was built with the ninfer-v3 recipe ptq1_gguf_mtp.py. The internal name is bonsai2-27b-uncensored. The converter tree is iamwavecut/ninfer-all at 63428ac249da8a6c5d891b508d7e483d8c7def6c.
6. The intermediate file was then repacked to T5. T5 keeps the same ternary integers and group scales, uses a different byte order, and allows INT8 activations. Stored weights stay the same, so generated text can differ from the PTQ1 pack. The packing comes from JGamboa/ninfer-4090-windows at e905787f067de89ef6ef28c52d65a2ce91d425ca.

Local check files: Bonsai2-PTQ1-MTP-r3-Q4-Uncensored.ninfer.conversion.json and Bonsai2-T5-MTP-r3-Q4-Uncensored.ninfer.repack.json.

Figures published upstream

These figures are from the OS-Software GGUF card. They describe that GGUF release, not this T5 repack.

They also report WikiText-2 perplexity of 10.1435 for the uncensored PTQ1_0 file, 16.7744 on a Japanese harmless sample, and 16/16 short functional checks, at a context of 512. Their card says long context and vision were not evaluated for that release.

Prism ML publishes these sampling settings for the original model's thinking mode: temperature 1.0, top_p 0.95, top_k 20, min_p 0.05. This directory has no separate measurement of whether the uncensored repack keeps the same behavior.

License

Along the weight lineage, Qwen, Prism ML, and OS-Software all declare Apache-2.0. Redistribution should keep the upstream attribution and license text.

The Heretic program itself is under the GNU Affero General Public License v3. That license applies to the program. The weight card declares Apache-2.0.

Citation

The Qwen team publishes this entry on the 27B model card:

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}

27B weights: https://huggingface.co/Qwen/Qwen3.8-27B

Ternary base:

@misc{prismml2026bonsai227b,
    title = {Ternary Bonsai 2 27B},
    author = {{Prism ML}},
    year = {2026},
    url = {https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf},
    note = {Documentation: https://docs.prismml.com/models/bonsai-2-27b}
}

Uncensored weights:

@misc{ossoftware2026bonsai2heretic,
    title = {Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF},
    author = {{OS-Software}},
    year = {2026},
    url = {https://huggingface.co/OS-Software/Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF},
    note = {Heretic low-rank edit written into the ternary codes; scales unchanged}
}

Citation supplied by the Heretic author:

@misc{heretic,
    author = {Weidmann, Philipp Emanuel},
    title = {Heretic: Training-free language model modification},
    year = {2025},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/p-e-w/heretic}}
}

The Heretic documentation points to this paper for the refusal-direction result:

@inproceedings{arditi2024refusal,
    title = {Refusal in Language Models Is Mediated by a Single Direction},
    author = {Arditi, Andy and Obeso, Oscar and Syed, Aaquib and Paleka, Daniel and Panickssery, Nina and Gurnee, Wes and Nanda, Neel},
    booktitle = {Advances in Neural Information Processing Systems},
    year = {2024},
    url = {https://arxiv.org/abs/2406.11717}
}

Packing and conversion:

• Converter tree: https://github.com/iamwavecut/ninfer-all/tree/63428ac249da8a6c5d891b508d7e483d8c7def6c
• T5 packing source: https://github.com/JGamboa/ninfer-4090-windows/tree/e905787f067de89ef6ef28c52d65a2ce91d425ca