Baseten announced a partnership and standard through its research arm Base Labs with Hugging Face and Goodfire AI to develop safety evaluation and monitoring infrastructure for open-weight models. The announcement dated September 17, 2026, aims to embed safety measures into the model training and deployment processes from the outset, rather than adding them later.
Open-weight models are under scrutiny because of a technique called “abliteration,” which enables the removal of safety measures. More than 6,000 models with abliteration applied are currently listed on Hugging Face. Base Labs will develop and publish methods for training and monitoring open models.
The partnership’s technical operation was not disclosed. Goodfire AI, which examines how model decisions are made, stands out as the company that could play a role in the goal of embedding safety into models. Baseten, an AI inference service provider, raised a $1.5 billion Series F round in June, bringing its valuation to $13 billion. Goodfire AI, meanwhile, raised a $150 million Series B round led by B Capital earlier this year.
Baseten issued an open call to the developer ecosystem to contribute to the development of the framework. The company said it aims to build an ecosystem in which open models are safe and accessible to everyone.
Why it matters
The ability to remove safeguards in open-weight models shows that measures taken after a model is released may not be sufficient on their own. For this reason, it is becoming important to determine at what stage and according to which criteria safety assessments will be conducted for developers, organizations using the model, and users accessing the open-model ecosystem. The announced approach aims to make safety part of the model development process without giving up accessibility; however, it is not yet clear how this will be applied to monitoring existing models. Since the partnership’s technical operation has not been disclosed, questions remain over how the standards will classify models with abliteration applied, measure risks, and incorporate developer contributions into the process.
Background
Face is not a new name in the FikirPilot archive: we have published 8 articles mentioning this name in the last 90 days; the most recent is dated September 13, 2026.
Term: open weight
Open weight means that the trained parameters of an AI model are downloadable; it does not mean the same as open source, as the training data and code may not be open.