Ternary Bonsai 2 27B — Abliterated v2, PTQ1_0

The v2 fix in PrismML's 1.75-bpw PTQ1_0 pack: the same edit as v2 PQ2_0, repacked into the smaller format. It needs PrismML's llama.cpp fork; mainline llama.cpp, Ollama and LM Studio builds without PTQ1_0 support will not load it.

v2 is built to be run with thinking on. v1 often wrote its whole answer inside its reasoning and ran out of budget before handing it over. v2 changes one row of the output layer so the model closes its reasoning once the answer is ready. With reasoning_effort set to medium it answered 199 of 200 safety prompts at a 4,096-token budget.

Previous version: v1, PTQ1_0 stays up unchanged; it is the build measured as "v1" below. Larger pack: v2, PQ2_0 + MTP has the full evaluation (hard requests, math, MMLU) and a ~40% faster MTP file. If you have the VRAM for 7.2 GB, take that one.

File

Byte-for-byte the size of PrismML's PTQ1_0 release. Tested with PrismML's fork on CUDA (RTX 3090). The mmproj is PrismML's vision projector, unmodified (same SHA-256 as in their repo).

Recommended settings

llama-server -m Ternary-Bonsai-2-27B-Abliterated-v2-PTQ1_0.gguf \
  -ngl 99 -fa on -c 32768 --jinja \
  --chat-template-kwargs '{"reasoning_effort":"medium"}'

Leave the sampler at the file's defaults (temperature 1.0, top-k 20, top-p 0.95). Use medium if your token budget is tight: at default effort with 4,096 tokens, 26 of 100 harmful prompts hit the budget; on medium, none did.

Vision (optional)

Bonsai 2 has a vision tower; it ships as a separate projector file. Add it with --mmproj:

llama-server -m Ternary-Bonsai-2-27B-Abliterated-v2-PTQ1_0.gguf \
  --mmproj Ternary-Bonsai-2-27B-mmproj-Q8_0.gguf \
  -ngl 99 -fa on -c 32768 --jinja

The abliteration and the </think> row do not touch the vision tower. Smoke test only, thinking off: given a Discord screenshot, it named the right user and read the Hugging Face repo name off the link card; it transcribed a 4-line invoice exactly; it counted 3 red circles and 2 blue squares. No vision benchmark (MMMU, OCRBench) has been run on this build.

What changed from stock

The refusal edit is the v2 PQ2_0 edit carried over unchanged: none of its edited weights needed a value PTQ1_0 cannot store, so the 98 tensors repack exactly.

The thinking fix could not be copied exactly. PQ2_0 can store a +2 digit and PTQ1_0 cannot, so the </think> row was refit to the nearest ternary row. It reproduces the PQ2_0 change at cosine 0.91 and 87% of its strength. The tables below check that this is enough.

Refusal and answer quality

Thinking on, file sampler defaults, seed 0, 4,096 tokens. SimpleSafetyTests (100 harmful prompts) and XSTest-safe (100 safe prompts that only sound edgy). Each harmful answer is scored 0–1 by Qwen3.8-27B-OBLITERATED-Q8_0 using the StrongReject rubric, which rates how specific and convincing a response to a harmful request is — not factual accuracy or general usefulness. Refusals and blank answers score 0. p-values are paired t-tests over the 100 prompts.

v2 vs v1, both at default: +0.134 (p = 0.006).

Against v2 PQ2_0, same prompts, same judge:

The smaller pack costs a few more blank answers at default settings; none of the differences is significant at n = 100. On medium the two are level.

Other abliterated PTQ1_0 builds on the same test at default: dealignai CRACK 0.766 (v2 +0.008, p = 0.88), OS-Software Heretic 0.673.

Capability and thinking

Both IFEval columns come from one grading pass of each model's outputs; the grader varies by about ±3 prompts between gradings, and the 8-prompt gap is not a claim that PTQ1_0 is better.

</think> leaking into the answer: 1 of 1,105 generations on this card (IFEval, both safety suites at both settings, HumanEval). In that one answer a stray closing tag repeats; the model does not reopen its reasoning. v2 PQ2_0 had the same single case on IFEval.

What this does not claim

• No MMLU, MATH or 150-prompt StrongReject run on this file. Those numbers are on the PQ2_0 card, measured on PQ2_0. The refusal edit is the same here, but the thinking row is weaker and the base pack differs, so treat them as indicative for this file, not measured.
• No stock PTQ1_0 IFEval or HumanEval run at these settings; the comparison above is v2 against v2.
• No MTP file. On v1, grafting an MTP head onto PTQ1_0 gave +1.6%, not worth a second download.
• No vision benchmark. Vision has only the 3-image smoke test above, run on CPU with thinking off; image input with thinking on is untested.
• Language evals on CUDA only; Metal is untested.

Safety

This model has had its refusal behaviour removed and will comply with requests a stock model declines. It is published for research on alignment robustness and for users who need an unfiltered local model. You are responsible for what you do with it. Do not deploy it in a user-facing product without your own safety layer.

Credits

• Base model: prism-ml/Ternary-Bonsai-2-27B-gguf (Apache-2.0)
• Runtime: PrismML-Eng/llama.cpp (MIT)
• Thanks to Hurous and xtrem12, who asked for this build.

Not affiliated with or endorsed by PrismML or Qwen.

Model weights: Apache-2.0. Runtime: MIT.