Base Labs partners with Hugging Face and Goodfire on open-weight model safety

Base Labs, the research arm Baseten launched earlier this year, announced Wednesday a partnership with Hugging Face and Goodfire AI to build evaluation and monitoring infrastructure for open-weight models. The move comes amid a sustained debate over the risk profile of such models, which can become dangerous when their safeguards are stripped away through a technique known as abliteration — a method that has gained momentum in recent months.

Hugging Face data shows its platform now hosts more than 6,000 abliterated models, illustrating both how easily the procedure can be carried out and the scale of potential exposure. Any developer with access to the weights can, in theory, remove the alignment mechanisms and distribute an unrestricted version. Baseten frames the initiative as a “standard” for open models — one embedded in the training and deployment process itself, not bolted on after the fact.

The company argues that openness is a safety advantage: it provides greater visibility into model behavior and lets safety research become transparent, actionable controls, unlike closed models where the mechanisms remain hidden. Base Labs will develop and publish methods for training and monitoring open models, with a focus on transparency and early integration. Technical details of the partnership have not been disclosed, though Goodfire — which specializes in opening the “black box” of models and explaining their decision-making — is the natural fit for the embedded component.

Both partners bring substantial resources. Baseten, an AI inference provider, raised a $1.5 billion Series F in June that valued the company at $13 billion. Goodfire secured a $150 million Series B earlier this year, led by B Capital, to advance its interpretability platform. The funding increases the odds that the proposed framework will reach practical implementation rather than remain a statement of intent.

Baseten has issued an open call for developers to contribute to the emerging framework. “Together, we’re building an ecosystem of open models that are safe and accessible to everyone,” the company said. The central question remains whether a voluntary standard can stem the flood of unguarded models already populating public repositories, or whether the effort amounts to regulating a reality that has already slipped beyond control.