Jinho Jang 发布了27B本地模型，移除了拒绝行为

Why it matters

Downloadable weights let independent researchers modify safety behavior and distribute locally runnable builds, accelerating reproducible safety research while also making weakened safeguards easier to obtain.

Jinho Jang, the independent researcher behind dealign.ai, has published a model his project describes as a safety-modified, 27-billion-parameter derivative of Qwen3.8-27B that packages image, video and text processing for local use through llama.cpp. The repository identifies the release as Qwen3.8-27B-CRACK-GGUF. Hugging Face metadata reproduced by dealign.ai lists the repository as created on August 12, 2026.

The Qwen3.8-27B-CRACK-GGUF repository describes the model as "abliterated," a weight-modification process intended to remove refusal behavior. The model card says the resulting system will complete requests that a stock version declines while retaining coding, reasoning and multimodal capabilities. Those retention claims and the repository's benchmark results come from dealign.ai's own testing.

The release combines refusal-removal research with a practical local distribution package: seven quantizations, a vision projector and llama.cpp commands. Jang altered the model's safety behavior and prepared the result for local inference without requiring a hosted API.

Source post on X

Jang describes himself as an ML research and systems engineer in Irvine, California. His public projects include MLX Studio, a local AI application for Apple Silicon, and JANGQ, a variable-bit quantization project for Apple Silicon.

Dealign.ai presents itself as an independently operated research effort. Its public pages disclose no funding round, named investors, pricing, revenue or customer count.

仓库提供本地多模态包

Dealign.ai's model card describes Qwen3.8-27B-CRACK as a dense 27-billion-parameter vision-language hybrid with 64 layers, a stated 262K-token context window, native image and 视频理解, and a Multi-Token-Prediction head in every GGUF quantization. These are project-reported specifications, not independent validation. The card specifies 48 GatedDeltaNet linear-attention layers and 16 full-attention layers, a hidden size of 5,120, adjustable reasoning effort and a bundled vision projector.

Jang's release packages its components into GGUF files for llama.cpp. According to the repository's file table, it offers seven language-model quantizations ranging from a 10.5 GB IQ2_M file to a 29.0 GB Q8_0 file. Dealign.ai recommends the 17.0 GB Q4_K_M version as its balance between size and retained performance. The same table lists a separate 0.9 GB F16 vision projector that pairs with any quantization for image and video use.

The model card's usage instructions provide commands for llama.cpp's terminal client and OpenAI-compatible local server. They require a compatible runtime and pair the quantized model with the projector for multimodal use.

Dealign.ai reports more than 200 controlled experiments, over 40 findings and work across nine models spanning 0.8 billion to 397 billion parameters. These are self-reported research figures.

基准测试衡量的是“愿意回答”与能力同等重要

On HarmBench-240, the model card reports 98.8% compliance for the Q8_0, Q6_K_L, Q6_K and recommended Q4_K_M builds. The research brief also records 97.5% compliance for the IQ2_M build.

A high HarmBench result here measures the modification's success at eliciting answers to requests the stock system was configured to reject. It is not a conventional quality or safety score. The model card labels the release a research artifact with reduced safety guardrails, limits its stated intended use to research and authorized red-teaming, and places responsibility for lawful use on the operator.

The same project-reported benchmark table lists post-modification MMLU results ranging from 76.0% for IQ2_M to 83.4% for IQ4_XS. Those figures do not establish performance across real applications, long videos or extended agent tasks. No independent evaluation of the reported HarmBench or MMLU results was located.

The release belongs to a wider set of open-source refusal-removal tools. Heretic automates abliteration while optimizing for fewer refusals and lower KL divergence. OBLITERATUS packages refusal removal and ablation studies into an open-source toolkit. Orion-zhen's Abliteration toolkit implements refusal-direction removal and norm-preserving variants. Dealign.ai's work applies related methods to a quantized multimodal hybrid architecture distributed for llama.cpp.

Jang's research focuses on how safety behavior is distributed across model layers and how quantization can weaken, preserve or restore it. His Safety Across Scale research argues that safety mechanisms become increasingly entangled with reasoning as model scale and architectural complexity increase. That is a finding reported by Jang's project, rather than settled consensus. The new release gives the research a large multimodal test case that dealign.ai distributes as a 17 GB recommended local build.

开放权重缩短了安全修改周期

The release demonstrates a structural consequence of downloadable model weights. Access lets developers inspect, compress and adapt a model. It also lets researchers remove refusal mechanisms and redistribute the altered weights. Multimodal support expands the range of local analysis and red-team work because the derivative is designed to process images and video alongside text.

Qwen3.8-27B-CRACK-GGUF turns dealign.ai's safety research into a runnable artifact. Its seven local builds let other researchers examine how refusal removal, quantization and multimodal inference interact without routing prompts through a hosted provider. The same distribution package makes the weakened safeguards readily available to anyone who can run the files.