Jinho Jang Puts a 27B Refusal-Removed Model Into a Local Download
The release turns refusal-removal research into a practical local package for text, images and video. Its capabilities and benchmark results remain project-reported, while the weakened guardrails are available without a hosted API.
Story brief
3 key pointsDealign.ai has released downloadable GGUF weights for Qwen3.8-27B-CRACK-GGUF, a 27B multimodal derivative whose weights were modified to reduce the base model’s refusal behavior. The package runs through llama.cpp rather than a hosted API, with language-model files spanning 10.5GB to 29GB plus a 0.9GB vision projector. Dealign.ai reports 98.8% HarmBench-240 compliance on several larger quantizations, but those...
- 01
The recommended Q4_K_M build is 17.0GB; the smallest IQ2_M is 10.5GB and the largest Q8_0 is 29.0GB.
- 02
A separate 0.9GB F16 vision projector supports image and video inference across quantizations.
- 03
The derivative claims a 262K-token context window, but extended-task and long-video performance remains unvalidated.
Jinho Jang’s dealign.ai has published Qwen3.8-27B-CRACK-GGUF, a 27-billion-parameter derivative designed to run locally after a weight modification intended to remove the refusal behavior found in the stock model. The release makes a reduced-guardrail system available as a multimodal package rather than a hosted service.
The project calls the modification “abliterated.” In the model card’s description, the goal is to make the system complete requests that the original model would decline. That is a change to model weights, not a prompt setting or a switch applied by an API provider.
A package built for local operation
The distribution is arranged for llama.cpp, an environment for running models locally. It includes commands for a terminal client and an OpenAI-compatible local server, so an operator can use the files without a hosted API. Repository metadata lists an August 12, 2026 creation date.
What the download contains
- Seven GGUF language-model quantizations, ranging from a 10.5 GB IQ2_M file to a 29.0 GB Q8_0 file.
- A separate 0.9 GB F16 vision projector that is intended to pair with any quantization for image and video work.
- A 17.0 GB Q4_K_M build that dealign.ai recommends as its balance of file size and retained performance.
- A stated 262K-token context window for the derivative.
The IQ2_M quantization is listed at 10.5 GB.
Dealign.ai identifies Q4_K_M as its recommended balance of size and retained performance.
The Q8_0 quantization is listed at 29.0 GB.
A benchmark can measure willingness, not safety
The project reports 98.8% compliance on HarmBench-240 for its Q8_0, Q6_K_L, Q6_K and recommended Q4_K_M builds. In this setting, the score measures whether the modification elicits answers to requests that the original model was configured to reject; it is not a conventional quality or safety score.
Dealign.ai also reports post-modification MMLU results from 76.0% for IQ2_M to 83.4% for IQ4_XS. Those results, along with the HarmBench figures, context claim and other specifications, are project-reported rather than independently validated. They do not establish performance in real applications, long-video work or extended agent tasks.
The distribution tradeoff
The model card labels the release a research artifact with reduced safety guardrails. It limits intended use to research and authorized red-teaming, and assigns responsibility for lawful use to the operator. Those terms frame the release, but the same local distribution package makes the altered weights available to people able to run them.
That creates a useful test case for researchers studying how refusal removal interacts with quantization and multimodal inference. It also puts the operational consequence of downloadable weights in plain view: the same access that supports inspection and adaptation can support redistribution of a model whose refusal behavior has been deliberately weakened.
Editorial analysis
Our Read
The notable shift is not merely that refusal removal is possible. It is that this release packages it into a local, multimodal setup with several size options and a recommended 17 GB build. That lowers the gap between a research intervention and routine deployment by an operator. The next evidence worth watching is independent testing of both sides of the tradeoff: whether the reported capability retention holds across the quantized builds, and how the system behaves on image and video inputs after its guardrails have been reduced. The model card’s research-only framing does not technically restrict who can run downloaded weights.