license: apache-2.0
language:
- en
pipeline_tag: text-generation
library_name: llama.cpp
base_model: prism-ml/Ternary-Bonsai-2-27B-gguf
base_model_relation: finetune
tags: - bonsai
- ternary
- gguf
- 2-bit
- prismml
- abliterated
- mtp
- android
- mobile
- on-device
- tokforge
Published by TokForge for on-device inference.
TokForge
- Website: https://tokforge.ai
- Discord: https://discord.gg/Acv3CBtfVm
- Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
- iOS TestFlight: https://testflight.apple.com/join/jnufjzRr
Runs on-device in the TokForge app.
TokForge-Bonsai-2-27B-Abliterated-GGUF
Bonsai 2 27B ternary, Abliterated. Prism ML's ternary Bonsai 2 27B in the PQ2_0 packing, with an MTP head, as abliterated by BoldingBuilds in their v2 release. This repo is an unmodified copy of that file, published so TokForge can pin it.
Created using Bonsai by Prism ML. The model and its ternary packing are Prism ML's (
prism-ml/Ternary-Bonsai-2-27B-gguf); the base model is Qwen3.8-27B by Alibaba Cloud; the abliteration is BoldingBuilds' (BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-GGUF, commite25d197aa62ce0a2f685fc65d76a41b2416c5e66). TokForge changed nothing in the file.
What it is
- Model: Ternary Bonsai 2 27B by Prism ML, built from Qwen3.8-27B.
- Packing: PQ2_0, Prism ML's 2-bit format for ternary weights (2.13 bits per weight).
- Abliterated: BoldingBuilds edited the refusal behaviour out of the released ternary weights in place, and changed one row of the output layer so the model closes its reasoning once the answer is written. It answers prompts that the stock Bonsai 2 27B refuses, and you are responsible for how you use it.
- MTP head: one multi-token prediction block (
blk.64), the MTP layer of Qwen3.8-27B. TokForge can use it to draft a few tokens ahead and have the full model check them in one pass; a drafted token is kept only when the full model would have produced it.
Using it in TokForge
- Needs a TokForge build from the 1.3.9 line or newer (the current Play beta, 1.3.8.1, does not list this model) and a phone with 16 GB of RAM or more.
- Download the file from this page and import it in TokForge.
- TokForge uses the MTP head on the CPU route only, and only on phones where its own check with and without the head measures a gain; elsewhere the file runs as a plain model.
- Other runtimes: PQ2_0 needs Prism ML's llama.cpp fork; llama.cpp builds without the PQ2_0 type cannot load it.
Provenance
| Source | BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP-GGUF at commit e25d197aa62ce0a2f685fc65d76a41b2416c5e66 |
| File | Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP.gguf, 7,657,489,696 bytes |
| SHA-256 | a4e4c7b578131595c1694354bd6c74d00920df1f5082647a6126899753ebebf8, the same as the source's file and its SHA256SUMS |
| Changes by TokForge | none: byte-identical copy, same file name, no header edit |
| Stock reference | Ternary-Bonsai-2-27B-PQ2_0.gguf from prism-ml/Ternary-Bonsai-2-27B-gguf, SHA-256 3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1 |
| What differs from stock | 99 of the 851 model tensors: ffn_down in blocks 15 to 63 (49 tensors), ssm_out in the linear-attention blocks 16 to 62 (36), attn_output in the full-attention blocks 15 to 63 (13), and one row of output.weight (the </think> token). 68,121,562 bytes in all, edited in place: every tensor keeps its type, shape and size |
| Unchanged from stock | the other 752 tensors (the token embedding, every norm, the attention input projections attn_q, attn_k, attn_v, attn_qkv and attn_gate, ffn_gate, ffn_up, the other ssm_* tensors, and everything in blocks 0 to 14), every other row of output.weight, and every header key except the two below |
| MTP head | blk.64, 15 tensors (Q8_0 weight matrices, F32 norms), 451,319,808 bytes, after the model tensors. Byte-identical to the head in our TokForge-Bonsai-2-27B-GGUF file Ternary-Bonsai-2-27B-PTQ1_0-MTP.gguf, which TokForge built from the mtp.* tensors of Qwen/Qwen3.8-27B at commit 1d4bf0f2ff. Not retrained on Bonsai |
| Header | qwen35.block_count 64 to 65 and qwen35.nextn_predict_layers = 1; no provenance keys, so this card is the record |
| Without the head | removing blk.64 and the nextn key gives BoldingBuilds' plain v2 file exactly (SHA-256 b284cbc6cb6c2894eb3d4181d805b966948128b2add9627dbd7d1e691724b3bb) |
BoldingBuilds' own description and measurements of the edit are on their card.
Files
| File | Bytes | SHA-256 |
|---|---|---|
Ternary-Bonsai-2-27B-Abliterated-v2-PQ2_0-MTP.gguf |
7,657,489,696 | a4e4c7b578131595c1694354bd6c74d00920df1f5082647a6126899753ebebf8 |
SHA256SUMS |
sha256 of the GGUF file | |
LICENSE, NOTICE.txt |
Apache-2.0 license and Prism ML's notice, copied unchanged from prism-ml/Ternary-Bonsai-2-27B-gguf |
Credits
- Ternary Bonsai 2 27B GGUF by Prism ML (Hugging Face). Used under the Apache-2.0 license. Created using Bonsai by Prism ML.
- Qwen3.8-27B by the Qwen team (Alibaba Cloud) (Hugging Face). Used under the Apache-2.0 license. The MTP head is its multi-token prediction layer.
- Abliteration v2 and the MTP file by BoldingBuilds (Hugging Face, commit
e25d197aa62ce0a2f685fc65d76a41b2416c5e66). Released under the Apache-2.0 license. Their card credits the MTP recipe to decent-jawfish, ProCreations and sudoingX/qwen38-mtp, and the head tounsloth/Qwen3.8-27B-GGUF. - Ternary kernels from Prism ML's llama.cpp fork (MIT); TokForge ported them to Android.
Not affiliated with or endorsed by Prism ML, Qwen or BoldingBuilds.
License
Apache-2.0, inherited from Prism ML's release, the Qwen base model and BoldingBuilds' release. The source repo ships no license file, so LICENSE and NOTICE.txt are carried unchanged from Prism ML's release. Per Prism ML's notice: created using Bonsai by Prism ML.