The Harrowgate Courier
Technology · Open source

The open-source model that nobody wanted to host

A widely praised language model sat unused for months because running it legally turned out to be the hard part.

A robotic hand against a network of lights

Free as in puppy: the weights were open, the responsibilities were not.

The release was triumphant: state-of-the-art weights, a permissive licence, benchmarks to make the giants blink. Then came the quiet question nobody had modelled — who actually runs this thing? — and the silence lasted four months.

Hosting a frontier model means answering for its outputs: the defamation, the medical advice, the homework, the scams. The licence gave away the weights; it could not give away the liability. Universities balked, startups blanched, and the model sat on a download page like a grand piano nobody could insure.

The hosts who said yes

The breakthrough came from an unlikely consortium: a national library, two universities and a cooperative cloud provider, splitting the legal risk the way they already split the compute. Their terms are strict — logged access, use policies with teeth, a kill switch each member can pull — and the queue to use the model is now six weeks long.

“Open weights are a gift. Running them is a job — and jobs need job descriptions.”

A template for the next one

The consortium has published everything: the hosting agreement, the acceptable-use policy, the incident log. Three more models have found hosts on the same terms. Open source, it turns out, was never the end of the story — it was the part where the community has to grow up.

Key points

  • Praised open model sat unhosted for four months over liability
  • Library-university-cloud consortium split risk and compute
  • Hosting template reused by three subsequent releases

The model's creator, asked what she learned, doesn't hesitate: “Release the weights and the warranty together next time.” The piano, at last, has somewhere to live — and a very thorough insurance policy.

The Dispatch

Signals, weekly, in plain English.

The technology stories that will matter to public life — without the hype cycle.