A problem that is easy to state
Base Labs, the research group Baseten spun up earlier this year, said on Thursday it will work with Hugging Face and Goodfire to develop and publish methods for training and monitoring open-weight models.
The immediate target is abliteration: the practice of stripping a model’s safety guardrails after release, so that it answers what it was trained to refuse. Hugging Face currently lists more than 6,000 abliterated models on its platform. Because the weights are public, this is not a vulnerability anyone can patch. It is a property of how open release works.
Why these three
The partners occupy three different points in the chain an open model passes through. Hugging Face is where the weights are distributed. Baseten is where a large number of them are served — it became an official Hugging Face Inference Provider in August 2026, so the plumbing between the two already exists. Goodfire works on interpretability, which is the closest anyone has to a way of inspecting what a model has actually learned rather than what it outputs.
“We believe openness to be an advantage for AI safety,” Base Labs said. “Openness provides more visibility.” Goodfire framed its own role more narrowly: “Safety must be built into open models and provided by those who serve them.”
That second sentence is the more interesting claim. It puts responsibility on the serving layer rather than the releasing lab, which is a different allocation from the one most open-weight debates assume.
What has not been announced
No technical detail has been published. The companies have not said what the methods will be, what would count as compliance with the standard they describe, or who would verify it. Base Labs is inviting developers from the wider ecosystem to contribute to the framework, which is the usual shape of an announcement made before the work exists.
Both companies have been raising heavily. TechCrunch reported that Baseten raised $1.5bn in a Series F in June 2026 at a $13bn valuation, and that Goodfire raised $150m in a Series B earlier in the year.
The test it will face
The hard part is not writing a standard; it is that abliteration takes a few hours on consumer hardware and the result can be uploaded anywhere. A method for training models that resist it, if one exists, would be a genuine research result rather than a policy. A method for monitoring what served models are doing is more tractable, but it only covers models served by participants — which is precisely not where an abliterated model ends up.
What would make this more than a press release is a published technique with an evaluation attached. That is what to look for next.
