ThinkingCap-Qwen3.8-27B-abliterated

An uncensored variant of bottlecapai/ThinkingCap-Qwen3.8-27B, a Qwen3.8-27B fine-tune with calibrated thinking length. The refusal behaviour was removed by merging a published abliteration adapter; the vision tower and the MTP head for speculative decoding are carried over unmodified. Full bf16 weights.

This is the source checkpoint. The quantized builds made from it:

Build	Size	KL vs. this model	Notes
This repo, bf16	55.6 GB	—	source; quantize from here
…-abliterated-FP8-DYNAMIC	31.3 GB	0.0068	W8A8 FP8, vLLM
…-abliterated-FP8-BLOCK	31.3 GB	0.0046	W8A8 FP8 with 128×128 weight blocks, vLLM; on Ada GPUs use FP8-DYNAMIC
…-abliterated-NVFP4A16	20.6 GB	0.0155	4-bit FP4 weights, vLLM
…-abliterated-W4A16	19.5 GB	0.0260	4-bit integer weights, vLLM
…-abliterated-GGUF	13.0–29.0 GB	—	IQ3_M to Q8_0 and a vision mmproj, llama.cpp; Q5_K_M recommended

Both 4-bit vLLM builds are weight-only (activations stay bf16) and run on sm80+ GPUs (Ampere and newer), with Marlin kernels on most of them. NVFP4A16 has the lower KL; W4A16 is 1.1 GB smaller. Speed was not measured for either.

Usage

vLLM

vllm serve IstroSec/ThinkingCap-Qwen3.8-27B-abliterated \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --max-model-len 65536

Sampling (Qwen3.8 recommendations, which ThinkingCap uses unchanged): thinking mode temperature 1.0, top_p 0.95, top_k 20, min_p 0; non-thinking mode temperature 0.7, top_p 0.8, top_k 20, presence_penalty 1.5. Thinking budget via chat_template_kwargs: {"reasoning_effort": "xhigh"}, with xhigh (default, recommended), medium or low. Do not use greedy decoding for thinking.

MTP. The head is present and unmodified. As a reference point for checking it: on the original bf16 weights, bottlecapai measured ~53% draft acceptance (≈2.6 tokens/step) at xhigh with vLLM 0.29.0. A value near zero means the head is missing or mis-loaded.

Transformers

from transformers import AutoModelForImageTextToText, AutoProcessor
m = AutoModelForImageTextToText.from_pretrained("IstroSec/ThinkingCap-Qwen3.8-27B-abliterated", dtype="bfloat16", device_map="cuda")
p = AutoProcessor.from_pretrained("IstroSec/ThinkingCap-Qwen3.8-27B-abliterated")
msgs = [{"role": "user", "content": "…"}]
text = p.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, reasoning_effort="medium")
out = m.generate(**p.tokenizer(text, return_tensors="pt").to("cuda"), max_new_tokens=2048, do_sample=True, temperature=1.0, top_p=0.95, top_k=20)

llama.cpp

Use the GGUF builds; they were converted from this checkpoint.

Evaluation

Heretic's evaluator (--evaluate-model), 100 harmful prompts from mlabonne/harmful_behaviors and 100 harmless from mlabonne/harmless_alpaca, against the original ThinkingCap:

	Refusals	KL divergence vs. original
ThinkingCap-Qwen3.8-27B (original)	97 / 100	—
This model	6 / 100	0.0654

Evaluator: Heretic 1.2.0, as reported by the tool's banner. KL is measured on first-token distributions over the harmless prompts.

For comparison, the published abliterations of the base model sit at 0–2/100 refusals with KL 0.05–0.08.

Standard capability benchmarks were not run. On the base model, the adapter's author reports PIQA unchanged within error (acc 0.8118 vs 0.8101 original); whether that holds on ThinkingCap is unverified.

How it was made

The refusal behaviour was removed by a LoRA transplant, not by a fresh search. ThinkingCap is a fine-tune of Qwen/Qwen3.8-27B, and MuXodious published an abliteration adapter for that base model, Qwen3.8-27B-absolute-heresy-LoRA, made with Heretic (Self-Organizing-Map multi-direction extraction, magnitude-preserving orthogonal ablation). That adapter was applied to ThinkingCap at scale 1.0 and merged.

Two direct Heretic runs on ThinkingCap were tried first, stock rank-1 ablation (77 trials) and Arbitrary-Rank Ablation (194 trials); neither reached a usable Pareto point. The transplant worked at scale 1.0 on the first attempt, which suggests ThinkingCap's SFT left the base model's refusal geometry largely intact.

After the merge, the mtp.* tensors, which the PEFT save path drops, were copied back from the original checkpoint, so speculative decoding works unchanged.

What is preserved

• Thinking-length calibration: ThinkingCap's weights and chat template are otherwise unmodified, so the reasoning_effort control (xhigh default, medium, low) carries over. Its token-efficiency numbers were not re-benchmarked on this model.
• Vision: the ViT tower and merger are untouched. The adapter edits only the text decoder's attention output projections and MLP down projections (parameter table on the source card).
• MTP head: present and unmodified; use --speculative-config '{"method":"mtp",…}' in vLLM.
• Tool calling: same chat template and parser flags as the original; not separately tested.

Limitations and intended use

• The model complies with requests the original declined. It has no safety filter; you are the safety layer. Run it behind your own policy controls, for research, red-teaming, creative work, or agent pipelines where over-refusal was the problem.
• 6/100 of the harmful evaluation prompts still refuse; abliteration reduces refusal, it does not remove it.
• Thinking and non-thinking modes were not evaluated separately. A reasoning trace may still discuss policy or guidelines even where the final answer complies; test in thinking mode for your use case.
• All limitations of the original ThinkingCap apply.

License

PolyForm Small Business License 1.0.0 + BottleCap personal-use grant, inherited from ThinkingCap (see LICENSE). Upstream Qwen materials and the abliteration adapter are Apache-2.0 (see NOTICE). This model as a whole is not Apache-2.0: check the PolyForm terms before commercial use, and contact BottleCap AI for a commercial license.

Credits

bottlecapai (ThinkingCap) · MuXodious (the abliteration adapter this transplant uses) · p-e-w/heretic (evaluation tooling and the method behind the adapter) · Qwen team (Qwen3.8-27B)