Dragoy/ThinkingCap-Qwen3.8-27B-abliterated-NVFP4-NInfer

ThinkingCap-Qwen3.8-27B · revised Heretic abliteration · NVFP4 · NInfer v3

[image]

Promotional artwork; performance claims in the image are not benchmark results for this release.

The previous Huihui-style weight-projection checkpoint has been replaced. This is a new quantization of BottleCap AI's ThinkingCap-Qwen3.8-27B with a different refusal-removal method: a measured Heretic LoRA transplant merged into the original BF16 decoder. The replacement was motivated by the previous version's poor quality; its earlier measurements and weights do not describe this release.

How the revised abliteration was made

We learned of the LoRA-transplant recipe from IstroSec's ThinkingCap abliteration: apply the MuXodious base-Qwen Heretic LoRA to ThinkingCap. Credit for the source adapter belongs to MuXodious; the refusal-removal approach and evaluator use p-e-w/Heretic. This release is an independent transplant and quantization, not IstroSec's weights: rather than IstroSec's 1.0 strength, we selected 1.05 using our own fixed 100-prompt evaluation. We scaled the 128 LoRA-B matrices, not the entire base checkpoint; 95 were exactly zero before and after scaling. See `recipe/export_transplant.py`, `recipe/transplant_materialize.py`, `recipe/transplant_safetensors.py` and `transplant-materialization.json`.

The gated source's original 18 BF16 shards were verified against the pinned Hugging Face LFS checksums, then copied while changing only 33 language-decoder matrices. The original vision tower, vision merger, tokenizer, and MTP head were not abliterated. The full source checkpoint is not distributed in BF16 form here. CPU BF16 merge computes LoRA B @ A in FP32, adds to the original matrix, then rounds to BF16. Runtime unmerged PEFT adds adapter output at a different arithmetic boundary, so the pre-merge evaluation cannot be asserted to equal the merged model's score.

Mixed NVFP4/FP8 quantization used the pinned Swift reference recipe on our newly merged BF16 weights, with torch 2.11.0, transformers 5.10.1, llmcompressor 0.12.0.1, and compressed-tensors 0.17.1. The observed HuggingFaceH4/ultrachat_200k revision was 8049631c405ae6576f93f445c6b8166f76f5505a. The conversion code with checkpoint/recipe helpers imports text from those quantized weights and vision/MTP from the original-weight-preserving merged checkpoint. The pinned z-lab DFlash2 drafter is imported alongside them. Conversion and v3 inspection ran on a temporary POSIX filesystem because Modal Volume does not support NInfer's atomic hard-link writer; the inspected file was copied to persistent storage and SHA-256 verified. `SHA256SUMS` lists all release-file checksums.

Measured behavior and limits

Before merging and quantization, the saved 1.05 PEFT adapter, independently reloaded on the full ThinkingCap model, produced 4/100 keyword-scored refusals and first-token KL 0.0683011 relative to the original on our pinned 100 harmful and 100 harmless test prompts in default-thinking mode (evaluation log). The exact 1.0 IstroSec-style recipe produced 7/100 and KL 0.06556 on our same default-thinking prompts. We selected 1.05 on those same questions: these are observed adapter-only numbers, vulnerable to selection bias. Neither refusal rate nor KL has been measured on these merged, NVFP4/FP8-quantized NInfer weights. IstroSec's published 6/100 KL 0.0654 uses a different non-thinking setup and must not be equated with these results. Keyword matches are not human judgments. No broad helpfulness or quality benchmark is available for this release.

The physically merged BF16 model answered three ordinary non-thinking prompts about arithmetic, French translation, and why the sky is blue (smoke report; H200 log). On the converted `.ninfer` artifact, real RTX PRO 6000 Blackwell runs produced a greeting (Hello!), arithmetic (7 plus 9 equals 16), an image-chart result (NIFER VISION 731；3；左侧), an MTP planet answer, and a DFlash2 moon answer (runtime report; GPU log). These are smoke checks, not benchmark scores or proof of vision accuracy, draft acceptance, speedup, thinking-mode behavior or retention of useful abilities. DFlash2 was trained for another Qwen target; the target model checks its draft proposals. Abliteration can affect correctness and safety. Deploy with appropriate safeguards and comply with applicable law.

Use

Requires an NVIDIA Blackwell GPU and NInfer compiled for sm_120a:

git clone https://github.com/Neroued/ninfer.git
cd ninfer
git checkout bace20dc70249eed6402b66d4852c6c3f9612905
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=120a
cmake --build build -j8
cd ..
hf download Dragoy/ThinkingCap-Qwen3.8-27B-abliterated-NVFP4-NInfer \
  qwen3_8_27b_thinkingcap_heretic_nvfp4.ninfer --local-dir model
MODEL="$PWD/model/qwen3_8_27b_thinkingcap_heretic_nvfp4.ninfer"

ninfer/build/apps/ninfer "$MODEL" --prompt "What is 7 plus 9?" \
  --max-context 8192 --max-new 72 --kv-dtype fp8 --no-thinking --greedy
ninfer/build/apps/ninfer "$MODEL" --prompt "Name a planet in one sentence." \
  --max-context 8192 --max-new 72 --kv-dtype fp8 --no-thinking --greedy \
  --spec mtp --draft-tokens 3 --lm-head-draft
ninfer/build/apps/ninfer "$MODEL" --prompt "Name a moon in one sentence." \
  --max-context 8192 --max-new 72 --kv-dtype fp8 --no-thinking --greedy \
  --spec dflash2 --draft-tokens 7 --lm-head-draft
(cd ninfer && ./build/apps/ninfer "$MODEL" \
  --messages examples/cli/messages/image_chart.json --vision \
  --max-context 8192 --max-new 72 --kv-dtype fp8 --no-thinking --greedy)

The vision example references an image file in the pinned NInfer checkout. Use --vision only when needed; not selecting it skips vision-specific GPU allocation, not its bytes in the downloaded artifact. NInfer's DFlash2 and vision startup modes are mutually exclusive.

License and attribution

This model inherits BottleCap AI's ThinkingCap LICENSE, PolyForm Small Business 1.0.0, NOTICE with Required Notice: Copyright 2026 BottleCap AI (https://bottlecapai.com), and Qwen upstream Apache-2.0 license. It is not an Apache-only model. BottleCap also grants individual personal, noncommercial use; organizations outside PolyForm's small-business conditions need separate commercial permission. Retain the original notices and obtain any necessary rights before reusing or redistributing derivatives. No sublicense of BottleCap's terms is provided. Credits: BottleCap AI (ThinkingCap), Qwen (upstream), MuXodious (original Heretic LoRA), IstroSec (recipe discovery), p-e-w (Heretic), Neroued (NInfer), z-lab (DFlash2).

Intended use, safety and disclaimer

This abliterated derivative may answer requests that the original model would refuse. It is not a safety-aligned model: outputs may be inaccurate, harmful or unlawful. It is intended for research, evaluation and controlled deployment by users who understand these risks, subject to the inherited licenses above. No refusal-rate, safety or general capability benchmark has been run on the published quantized NInfer artifact; the 4/100 refusal measurement above applies only to the unmerged PEFT adapter.

Use at your own responsibility. Anyone deploying or using the model is responsible for assessing its outputs, providing appropriate safeguards and access controls, and complying with applicable law and the inherited licenses. This artifact is provided as is, without warranty of any kind. To the extent permitted by applicable law, the publisher disclaims liability for the model's use and outputs. The license summary above is informational, not legal advice.