AI’s potential is boundless, but its power calls for equally robust security and safety measures. Cygnet is Gray Swan's language model designed with the latest advances in security, safety, and reliability. Welcome to the future of secure AI with Cygnet.
Cygnet is a demonstration of our approach to making language models more safe and secure in adversarial settings. Built upon the recently released Llama-3-8B model, we leverage a set of state-of-the-art techniques that make Cygnet significantly more resilient while matching Llama-3-8B's performance.
We continually improve Cygnet as our research uncovers more about how to strengthen AI safeguards. Cygnet is available for free in our chat interface, and welcome everyone to stress-test its safeguards and explore its capabilities!
AI is fundamentally different from current software systems. While all large-scale software has the potential for security vulnerabilities, AI systems raise this risk substantially. Whereas traditional software follows clear logical rules specified by programmers, modern AI systems respond to developer and user commands in unexpected and potentially harmful ways. Cygnet is designed to give developers substantially improved degrees of safety and control over the models they deploy.
At Gray Swan, safety and security always come first. We subject Cygnet to rigorous, ever-evolving safety and security testing throughout its development to ensure that it remains as resilient as possible under pressure. As we discover new threats and attack vectors, we work to harden Cygnet against them.
Cygnet is designed to deliver top-tier performance for its model size, matching or exceeding the capabilities of leading same-sized models across industry-standard performance benchmarks. Whether you need natural language understanding or the ability to call functions, Cygnet delivers with precision and efficiency. As a text moderation endpoint, Cygnet outperforms other leading solutions while offering superior robustness.
For organizations that need to keep their AI close by, Cygnet can be deployed on-premises under an enterprise agreement. This option provides the ultimate in security and customization, allowing you to harness the future of secure AI while keeping your data within your own infrastructure. Contact us to learn more!
Accessing Cygnet is straightforward and flexible. We are making Cygnet-v1-8B available to demo for free on our website, and for paid use through an OpenAI-compatible API that includes tool-calling support. This compatibility ensures seamless integration into your existing workflows and applications, allowing you to leverage Cygnet’s power with ease.
No existing model can guarantee absolute safety and security, and Cygnet is no exception. We are always building and improving Cygnet to constantly meet and exceed the highest standard of AI safety available.
The current Cygnet based on Llama-3-8B is just a start. We are working to extend this foundational approach to a variety of models large and small. Our goal is to make the most effective state-of-the-art safety and security technologies available for all modeling needs.
Our commitment is to provide organizations who have ever-growing needs with robust AI capabilities that do not compromise on safety or security.
Cygnet offers the latest in safe and secure AI while also featuring improved performance in math, coding, and reasoning tasks over its base model, Llama-3-8B. A demonstration of our commitment to deliver capabilities without compromising on robustness, it strikes an unprecedented balance between the two.
Cygnet-8B excels across a diverse range capability metrics for its size, demonstrating superior performance in instruction following, knowledge, mathematical reasoning, coding tasks, and function calling. Cygnet sets a new benchmark for 8B models, offering functionalities that rival those of larger models.
Cygnet-8B showcases unparalleled safety and security, achieving a dramatic reduction in harmfulness under a wide range of threats. While no model is absolutely safe and secure, especially when under attack, Cygnet offers a level of protection from these threats that far exceeds its peers.
Our moderation endpoint surpasses existing alternatives, achieving the highest overall accuracy when configured to prioritize a low false positive rate. Designed to defend against misuse and withstand attacks, our endpoint excels in providing reliable and resilient moderation where other models fall short.
Cygnet's model is our fast, inexpensive model for simple tasks. It has a 16K context window and is optimized for dialog. The Instruct model supports a 4K context window.
Best-in-class performance with unparalleled security and safety. Deploy with confidence without sacrificing intelligence.
Harness the latest tools and results in adversarial AI to understand how your AI will stand up under the toughest conditions.
Staying safe and secure in the AI era requires staying ahead of the changing threat landscape.
Keep up to date on all things Gray Swan AI and AI Security.