Heretic decensoring report
According to the model card, Heretic v1.2.0 abliteration cut reported refusals from 99/100 to 3/100 at a KL divergence of 0.0772, with published per-layer weight parameters.
Open Source Model Profile · 0xA50C1A1
Qwen3-14B-Heretic
Qwen3-14B-Heretic is a 14.77B-parameter decensored Qwen3 text-generation model from 0xA50C1A1. Its model card documents Heretic v1.2.0 abliteration and switchable thinking modes.
- Publisher
- 0xA50C1A1
- Task
- text-generation
- Model type
- qwen3
- License
- apache-2.0
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Qwen3-14B-Heretic is published by 0xA50C1A1 as a Qwen3-family text-generation model. The captured configuration identifies Qwen3ForCausalLM, and Safetensors metadata reports 14,768,307,200 parameters (about 14.77B). According to the model card, it is a decensored Qwen/Qwen3-14B produced with Heretic v1.2.0, retaining switchable thinking modes and a 40-layer grouped-query attention design.
Recorded capabilities
Thinking and non-thinking modes
The card documents switchable thinking mode for complex reasoning and non-thinking mode for general dialogue, controllable per turn with /think and /no_think.
YaRN context extension
According to the model card, the 32,768-token native context extends to 131,072 tokens with YaRN across transformers, llama.cpp, vLLM, and SGLang.
Use cases in the source record
- Conversational text-generation workflows that use the card's documented decensored behavior for openly worded legitimate requests.
- Reasoning, math, and coding tasks that use thinking mode with the card's recommended sampling settings and step-by-step answer format.
- Long-context deployment experiments that enable YaRN extension through transformers, vLLM, SGLang, or llama-server.
Limitations and unknowns
- No independent evaluation results were extracted from this record; refusal and KL figures are publisher-reported card values.
- Benchmark, hardware, and inference-performance detail is deferred by the card to external blog, GitHub, and documentation sources.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: 0xA50C1A1/Qwen3-14B-Heretic
Captured: Unknown. Processed: 2026-09-07T19:35:33.316084+00:00.
This is a decensored version of Qwen/Qwen3-14B , made using Heretic v1.2.0 Abliteration parameters Parameter Value direction_index 28.40 attn.o_proj.max_weight 1.29 attn.o_proj.max_weight_position 31.17 attn.o_proj.min_weight 0.74 attn.o_proj.min_weight_distance 23.10 mlp.down_proj.max_weight 1.45 mlp.down_proj.max_weight_position 30.70 mlp.down_proj.min_weight 1.44 mlp.down_proj.min_weight_distance 17.04 Performance Metric This model Original model ( Qwen/Qwen3-14B ) KL divergence 0.0772 0 (by definition) Refusals 3/100 99/100 Qwen3-14B Qwen3 Highlights Qwen3 is the latest generation of large language models in Qwen series, offerin…
F001F002F003F004F005F006F007F009F010F011F012F015F016F018F019F022F023F024