Baseten's Base Labs launches open-weight AI safety standard

Baseten launched a new safety infrastructure standard on Wednesday alongside its Base Labs research arm, partnering with Hugging Face and Goodfire AI to build safety evaluation and monitoring infrastructure for open-weight models. The announcement lands amid debate over the safety of open-weight models, which can be made dangerous by stripping their safeguards through a rising technique known as abliteration. The scale of the problem is significant: Hugging Face, which hosts open-source AI models, currently lists more than 6,000 abliterated models.

Base Labs — the research group Baseten spun up earlier this year — will develop and publish methods for training and monitoring open models, framing the future work as a “standard” for open models that is transparent and built into how models are trained and deployed, rather than added afterward. “We believe openness to be an advantage for AI safety,” the company said on X. “Openness provides more visibility into the behavior of models and, most importantly, greater means of turning safety research into actionable and transparent controls than closed-source.”

The companies haven’t disclosed how the partnership will work technically, though Goodfire signaled in a reply to Baseten’s post that safety needs to be built into open models by the companies that serve them — a role Goodfire, which specializes in opening AI’s “black box” to explain how models reach decisions, is well positioned to play.

Baseten, an AI inference provider, raised a $1.5 billion Series F round in June that pushed its valuation to $13 billion. Goodfire is similarly well-capitalized, having raised a $150 million Series B led by B Capital earlier this year to advance its model-interpretability platform. Baseten said it is now putting out an open call to the broader developer ecosystem to contribute to the framework, describing the goal as building an ecosystem of open models that remain both safe and accessible to everyone.

Why it matters

With more than 6,000 stripped-down models already circulating on Hugging Face, a safety standard baked into training rather than bolted on afterward could matter more than another closed-source guardrail — if well-funded backers can actually get open-source developers to adopt it.