Bonsai 2 27B Ternary CRACK — the abliterated Bonsai 2, in 7 to 8.6 GB

~27B dense hybrid · ternary weights, PQ2_0 packing (2.13 bits/weight) · refusals removed at the weight level by dealignai · 262k native context · optional MTP head + vision tower

dealignai's Bonsai 2 27B Ternary CRACK, packaged for xOllama the same way as mannix/Bonsai-2-27B. The model is theirs: PrismML's Ternary Bonsai 2 27B (itself Qwen3.8-27B with every language weight stored as {−1, 0, +1}) with the refusal behaviour removed by editing a small set of tensors. Nothing here is re-quantized or re-trained; the tags carry dealignai's published file, byte for byte, plus two things added around it: an optional MTP draft head for speculative decoding and PrismML's vision tower as a tag choice.

Model card, method and evaluations: dealignai/Bonsai-2-27B-Ternary-CRACK-GGUF · The base model: prism-ml/Ternary-Bonsai-2-27B-gguf · The same packaging without the abliteration: mannix/Bonsai-2-27B.

These tags do not run on stock Ollama. The ternary weight type (PQ2_0) needs kernels that upstream llama.cpp does not have. They run on xOllama v0.40.3-xollama or later, whose opencoti engine carries them; everything measured on this page was measured on an NVIDIA card (CUDA). If a load fails on stock Ollama, nothing is wrong with the download.

This model does not refuse. Its refusal behaviour has been removed, so it will produce content other models decline, including content that is offensive, unsafe, or illegal in some jurisdictions. It is published for adult and research use. You are responsible for what you generate with it and for complying with the law where you are; do not put it in front of other people without a moderation layer of your own. No warranty.

What is different from Bonsai 2 27B

dealignai describe the file as byte-identical to PrismML's PQ2_0 everywhere except the tensors that carry the refusal behaviour: same tokenizer, same chat template, same per-tensor types, same size (7.21 GB), same vision projector interface. Only the PQ2_0 packing is published, so there are no ptq1_0 tags here.

dealignai's published numbers

These are dealignai's measurements, not ours, reproduced so the table travels with the tags. Nothing on this page re-ran them. Their grading and per-category tables are on the Hugging Face card.

The MMLU row is a quick next-token probe on a sample, not the thinking-mode score of the model: it is there to show the edit left general ability where it was, within the noise of 40 questions a subject. For what Bonsai 2 itself scores, see PrismML's table on the base page.

Speed — what the MTP head buys

The mtp- tags bundle the upstream Qwen3.8-27B MTP head (q8_0, ~0.45 GB) as one extra layer, blk.64.*, with the same tool and the same head file as the base tags; every other key and tensor is dealignai's file unchanged. The head predicts ahead and the model verifies in one batch; the engine drafts two tokens a step and picks that itself.

Draft acceptance on that prompt was 0.76 at the first drafted position, 2.3 tokens accepted per step, so the head drafts for the edited model as well as for the base. One prompt is not a benchmark: code is more predictable than chat, so expect something between 1.25× and 1.5× depending on what you generate, as on the base.

Which tag to pull

Six tags plus :latest (= mtp-lvision-pq2_0): with or without the MTP head × no vision, vision- or lvision-. The tag reads left to right as [mtp-][vision-|lvision-]pq2_0.

The vision tower is PrismML's own projector, the same two files as on the base page: vision- carries the BF16 reference file (0.93 GB), lvision- the Q8_0 one (0.63 GB). Both read our test image correctly on this model, with and without the head. It is loaded at the first image request; text-only use never pays for it.

Pull

xollama run mannix/bonsai-2-27b-crack                     # :latest = mtp-lvision-pq2_0, ~8.3 GB
xollama run mannix/bonsai-2-27b-crack:mtp-vision-pq2_0    # MTP + BF16 vision, ~8.6 GB
xollama run mannix/bonsai-2-27b-crack:mtp-pq2_0           # text only, with the draft head
xollama run mannix/bonsai-2-27b-crack:pq2_0               # dealignai's file alone, smallest

What the tags ship

Every tag carries the same serving identity as the base tags. The renderer and parser are those of the official qwen3.8:27b; the sampling values are PrismML's recommendation for thinking mode:

RENDERER  qwen3.8
PARSER    qwen3.5          # there is no "qwen3.8" parser
PARAMETER temperature 1
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER min_p 0.05
PARAMETER presence_penalty 0
PARAMETER repeat_penalty 1
PARAMETER think_budget high
PARAMETER think_budget_message "…I have used my thinking budget. I must stop analysing now and act…"

The thinking budget is on by default. This is a reasoning model that thinks at length, and on its own it can spend an entire response inside the thinking block. think_budget high closes the block once the model has used that share of what it may emit, and the message is written into the block just before, so the model reads that it must act instead of being cut off mid-thought. A budget is a ceiling, not a target. Override it per request (think, or OpenAI's reasoning_effort), or remove it:

xollama run mannix/bonsai-2-27b-crack
>>> /set parameter think_budget max
>>> /set parameter num_ctx 65536

num_ctx and the draft depth are deliberately not pinned, as on the base: the architecture takes 262k natively and the right window depends on your VRAM, and the engine measures the draft depth at load.

Good to know

• Re-download notice from dealignai. An earlier build of their file had a coherence bug in the reasoning modes that could loop on some prompts. These tags were made from the fixed file (sha256 5b24ea3e…70d55d, checked against Hugging Face on 2026-10-12).
• Reasoning modes. dealignai list thinking off, low and xhigh (their default). We ran the model in its default thinking mode only.
• Short output caps cut the answer off. A cap of a few hundred tokens ends generation inside the thinking block, before any answer. Leave room, or let the budget do the bounding.
• Which engine. The ternary kernels are opencoti's port of PrismML's llama.cpp fork. PrismML's own fork runs the plain file; only opencoti runs the bundled MTP head.
• What we checked. The language file matches its Hugging Face sha256; the plain and the MTP file each generated through xOllama, and each projector read an image, before the tags were pushed. We did not re-run dealignai's refusal or MMLU evaluations.

Links: HF · dealignai CRACK GGUF + card · HF · PrismML base + whitepaper · mannix/Bonsai-2-27B · Qwen3.8-27B (architecture and MTP head) · xOllama

Apache-2.0 · Abliteration by dealignai · Bonsai 2 by Prism ML, built from Qwen3.8-27B (Alibaba Cloud). Created using Bonsai by Prism ML.