Qwen3.8-27B-Heretic2.1.0
uncensored multimodal Qwen3.8
2 min readupdated
Creator symmetrisch released an uncensored version of the flagship compact model Qwen/Qwen3.8-27B (27.36 billion parameters), processed with Heretic v2.1.0. It handles the same tasks as the original—text generation, coding, agentic workflows, and image and video analysis—without strict safety filters. Unlike the base model, which refuses almost all sensitive prompts, this build answers them directly while keeping its original capabilities and reasoning logic intact. It is a solid fit for developers and hobbyists looking for a capable local VLM (vision-language model) without extra restrictions.
Key features
- Minimal refusals: The refusal rate dropped from 98/100 on the original Qwen3.8-27B down to 9/100.
- Surgical modification: Uses Arbitrary-Rank Ablation (ARA, targeted removal of refusal directions across model layers) on layers 6 through 60, keeping KL divergence (a metric measuring drift from the base model's behavior) at just 0.1237.
- Native multimodality: Processes text, complex STEM diagrams, documents, and videos up to an hour long.
- Flexible reasoning control: Thinking mode is enabled by default, and you can toggle it off or adjust its depth via
reasoning_effort. - Massive context: Features a native context window of 262,144 tokens, extendable up to 1,000,000 tokens.
What you can do
- Local agent and coding assistant: Run long, multi-step terminal tasks and generate code without false-positive censorship triggers on security or vulnerability analysis prompts.
- Image and document analysis: Parse diagrams, charts, and long video recordings in your own pipelines using Hugging Face Transformers, vLLM, or SGLang.
- Research and roleplay: Use it for scenarios where standard models refuse to respond due to overly strict filters.
Example
According to the model card, it works out of the box with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. On a benchmark suite of 100 adversarial prompts, the original Qwen3.8-27B refuses 98 times, whereas Qwen3.8-27B-Heretic2.1.0 answers 91 out of 100 prompts.
Read more
DetailsDecensoring metrics: 9/100 refusals (vs. 98/100 on the base model), KL divergence — 0.1237.
Base Qwen3.8-27B benchmarks: SWE-bench Pro — 61.7, Terminal Bench 2.1 — 73.0, CoWorkBench — 70.7.
Architecture: 64 layers (hybrid Gated DeltaNet and Gated Attention), 27.36B parameters.
Reproducibility: ARA ablation parameters are published in the repository's
reproduce folder.









Comments