Baseten partners on safety infrastructure for open-weight AI models

Baseten has launched a safety infrastructure initiative through its Base Labs research arm, partnering with Hugging Face and Goodfire AI to develop evaluation and monitoring methods for open-weight AI models. The companies announced the effort on Wednesday as concern grows that model safeguards can be removed through a technique known as abliteration.

The scale of that concern is visible on Hugging Face, which currently lists more than 6,000 abliterated models. Baseten says Base Labs will develop and publish methods for training and monitoring open models, with the stated ambition of establishing a transparent safety standard that is incorporated into training and deployment.

Safety controls for models with published weights

Open-weight models can be inspected, adapted and served by a broad range of developers. That openness also means safeguards introduced by the original model developer may not remain intact after modification. Abliteration has emerged as one route for removing those safeguards, placing greater importance on evaluation and monitoring throughout a model’s lifecycle.

Baseten argued that openness can support AI safety by giving researchers more visibility into model behaviour and more ways to convert safety research into transparent controls. The initiative is framed around making those controls part of the model and its serving environment, rather than attaching them after a system has already been built.

Roles remain to be defined

The partners have not disclosed the technical design of the collaboration or specified the tools that will be released. Goodfire AI said safety must be built into open models and provided by those who serve them. Goodfire specialises in model interpretability, seeking to explain how AI models reach decisions, an area that could be relevant to the partnership’s goal of embedded safety.

Baseten provides AI inference services and raised a $1.5 billion Series F in June, giving the company a $13 billion valuation. Goodfire raised a $150 million Series B led by B Capital earlier this year to advance its interpretability platform.

An open call for contributions

Baseten is inviting the wider developer ecosystem to contribute to the framework. The announcement positions the work as an ecosystem effort around open models that are both safe and accessible, rather than a closed implementation controlled by one provider.

For businesses building or serving open-weight models, the practical implication is to evaluate safety controls after fine-tuning and deployment, not just at model selection, and to define monitoring responsibilities before systems reach users.