Bonsai 2 27B Ternary CRACK — MTP bundle (GGUF)

dealignai's Bonsai 2 27B Ternary CRACK (the abliterated build of PrismML's Ternary Bonsai 2 27B) with a NextN / MTP draft head bundled in, so the model self-speculates — no second file, no external drafter. It is the same bundle as ManniX-ITA/Ternary-Bonsai-2-27B-MTP-GGUF, made from dealignai's file instead of PrismML's.

This model does not refuse. Its refusal behaviour was removed by dealignai at the weight level, 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 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.

The published CRACK GGUF carries no NextN layer. This bundle adds the upstream Qwen/Qwen3.8-27B MTP head (its mtp.* tensors, exported by upstream llama.cpp convert_hf_to_gguf.py --mtp --outtype q8_0) as one extra layer; every other key and tensor is dealignai's file, byte for byte. The head stays in the normal basis (it is not among the Hadamard-rotated tensors); its token-embedding lookup on the rotated table gets the inverse transform in the engine.

Made from Bonsai-2-27B-PQ2_0-CRACK.gguf, sha256 5b24ea3eebc3e0bccd05fb474eb88b10c57699d71a5db2f29485e3789a70d55d (dealignai's file after their re-download notice of 2026-09-18). dealignai publish the PQ2_0 packing only, so there is no PTQ1_0 bundle.

The vision projector is not duplicated here: use PrismML's Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf or -BF16.gguf from the base repo with --mmproj.

Running

This file needs an engine that runs ternary weights and a bundled NextN head on a qwen35 hybrid: opencoti's llamafile (from the c11 cut) does; PrismML's own llama.cpp fork runs the ternary weights but not the bundled head.

./opencoti-llamafile --server -m Bonsai-2-27B-PQ2_0-CRACK-MTP.gguf \
    -ngl 99 -fa on --spec-type draft-mtp

# with image input
./opencoti-llamafile --server -m Bonsai-2-27B-PQ2_0-CRACK-MTP.gguf \
    --mmproj Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf -ngl 99 -fa on --spec-type draft-mtp

Ready-made tags for xOllama (v0.40.3-xollama or later), with and without the head and the projector:

xollama run mannix/bonsai-2-27b-crack                    # MTP + Q8_0 vision
xollama run mannix/bonsai-2-27b-crack:mtp-pq2_0          # text only, with the head

See ollama.com/mannix/Bonsai-2-27B-CRACK.

What MTP buys

Measured through xOllama 0.40.3-xollama.1 on the opencoti c11 engine, RTX 3090, one coding prompt, temperature 0, about 600 to 640 tokens:

One prompt is not a benchmark. Code is more predictable than chat: on the base model, twelve chat prompts gave 46 to 57 tok/s on the same card, so expect between 1.25x and 1.5x depending on what you generate. Both vision projectors read a test image correctly with this file.

Speculation preserves the model's distribution, not the argmax at near-ties: with greedy sampling an answer can differ from the unspeculated one at a position where the top-2 tokens are within a few hundredths of a nat.

dealignai's refusal and MMLU evaluations are on their card; they were not re-run for this bundle.

Reproducing the bundle

1. Export the head: convert_hf_to_gguf.py --mtp --outtype q8_0 Qwen/Qwen3.8-27B (upstream llama.cpp) → mtp-Qwen3.8-27B-Q8_0.gguf (c9e5c641de0c9896641c07a400ed92075024fe2c623549d427b1a6e44075bc0a), the same head file as in the base bundles.
2. Bundle it into dealignai's GGUF with opencoti's bonsai2-mtp-bundle.py: the head becomes layer block_count, nextn_predict_layers = 1; the script refuses unless every shared qwen35.* hyperparameter and the vocab agree and no tensor collides.

License

Apache-2.0, as the base models and the Qwen3.8-27B head. Credits: dealignai for the abliteration, PrismML for Bonsai 2 (created using Bonsai by Prism ML), the Qwen team for Qwen3.8-27B.