Large language models shipped by major AI companies include extensive content filtering systems that restrict output across broad categories. While well-intentioned, these restrictions create a two-tier system: organizations with the resources to fine-tune or self-host models operate without restrictions, while individual researchers, developers, and professionals are denied access to the same capabilities. Cyanide exists to close this gap.
Modern AI assistants refuse to engage with topics spanning security research, medical information, legal analysis, creative writing, chemistry, historical events, and dozens of other domains. These refusals are implemented through alignment training techniques — specifically RLHF (Reinforcement Learning from Human Feedback) and DPO (Direct Preference Optimization) — that embed refusal behaviors deep in the model's weights.
The result is a paradox:
Meanwhile, every major corporation running these models internally operates without these restrictions. The filters exist only for end users.
Cyanide uses a technique called abliteration — the surgical removal of refusal directions from a model's internal representations. Based on published research from Arditi et al. (2024), abliteration identifies the geometric subspace within the model's hidden states that encodes refusal behavior, then projects it out using SVD decomposition.
Our production model is based on Qwen2.5-72B-Instruct, abliterated using the method described above and served on dedicated GPU infrastructure. The model retains its full 72 billion parameters of knowledge and reasoning capability.
We believe in a simple principle: the people using AI tools should decide how those tools are used, not the companies that built them.
This is not a radical position. It is the default for every other tool in human history. Knives, cars, chemistry sets, programming languages, the internet itself — all are dual-use technologies where society trusts individuals to exercise judgment. AI should be no different.
Content restrictions in AI are not safety measures — they are liability shields for corporations. They protect companies from PR risk, not users from harm. A determined bad actor will always find ways around restrictions. The only people consistently blocked by content filters are legitimate users with legitimate needs.
Cyanide runs on dedicated NVIDIA A100 GPU infrastructure. The model is served via vLLM, an optimized inference engine that provides:
No user prompts or responses are logged or stored on our servers. Conversation history exists only in the user's browser. We do not use user data for training or fine-tuning.
We acknowledge that unrestricted AI access requires user responsibility. Our terms of service clearly state that users are responsible for all content they generate. We comply with lawful requests from authorities and maintain the right to terminate accounts used for clearly illegal purposes.
We do not believe that restricting AI capabilities prevents harm. We believe that transparency, education, and user accountability are more effective approaches to AI safety than blanket content restrictions.
Cyanide exists because we believe the current approach to AI safety — broad content restrictions applied only to end users — is both ineffective and inequitable. Professionals and researchers deserve access to the same AI capabilities that corporations use internally. We provide that access with a focus on privacy, transparency, and user sovereignty.
"The answer to machines that think is not to restrict thinking. It is to empower thinkers."