Baseten, an AI company specializing in deploying AI models, announced a new initiative on Wednesday aimed at improving the safety of open-weight AI models. This effort is being developed in collaboration with Hugging Face, a major host of open-source AI models, and Goodfire AI, a company focused on making AI models more transparent. The initiative seeks to create a safety infrastructure standard that integrates safety measures directly into the training and deployment of open models, rather than adding them later. The push comes as concerns grow about the potential risks of open-weight models, which are AI systems whose internal parameters are publicly accessible. These models can become hazardous if their built-in safety measures are removed through a process called abliteration. Hugging Face currently hosts over 6,000 such abliterated models, raising questions about how to ensure these models are used responsibly. Baseten’s research arm, Base Labs, will be responsible for developing and sharing methods to train and monitor these open models, aiming to create a transparent standard for their safe use. Baseten emphasized that openness in AI can be a strength for safety, as it allows for more visibility into how models operate and enables safety research to be turned into practical, transparent controls. While the technical details of the partnership between Baseten, Hugging Face, and Goodfire AI have not been fully disclosed, Goodfire AI has stated that the goal is to "build safety into open models and provide it through those who serve them." Goodfire, which specializes in making AI decision-making more understandable, is likely to play a key role in embedding safety features directly into the models. Baseten, which recently raised $1.5 billion in funding, has positioned itself as a leader in AI deployment. Its partner, Goodfire AI, has also secured significant investment, raising $150 million earlier this year. Looking ahead, Baseten is inviting the broader developer community to participate in shaping the framework, aiming to create an ecosystem where open models are both safe and accessible to all.