Cambridge spinout sells frontier model firms can host themselves

Flower Labs, a spinout from Cambridge University, has launched a general-purpose model called Endeavor and is selling it under licence.

The company says Endeavor completed tasks at similar levels to OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5. Those are the two US firms' main models.

The claim is the company's own. No benchmark suite, test set or scoring method has been published alongside it.

The selling point is where the model sits. Customers can deploy Endeavor on their local IT systems, which Flower Labs says gives them reliable access, control and better security.

The leading US providers ship closed-source models reached through software interfaces. That access can be withdrawn.

It happened this year. In June the US government and Anthropic temporarily restricted access to Fable 5 on US national security grounds, and Downing Street asked for an exemption.

Endeavor is itself a closed-source model. What it is not is a rented interface.

It also supports federated learning, a training method that lets a model learn from sensitive data without that data moving to a central server. Flower Labs says healthcare, financial services and defence companies, plus governments, are interested for that reason.

Underneath, the product is an assembly job. Flower Labs draws on freely available models from companies including Meta and Mistral, adds the reasoning of its own model, Lizzy, and wraps the result in proprietary harnesses — the software that turns a model into a working agent.

Professor Nicholas Lane, 47, co-founder and chief scientist, teaches machine-learning systems at Cambridge and St John's College. He said the performance gap between closed-source and open-weight models had become "vanishingly narrow".

He founded the company in 2023 with Daniel Beutel and Taner Topal. They went through Y Combinator the same year, and have since raised $23.6 million.

Flower Labs reports that its models, agents and infrastructure have been used in more than 2,500 organisations, including JP Morgan, the IBM-owned Red Hat and the US Department of Energy. Those are the company's own figures.

Lane said the long-term plan is to build more of its own proprietary models, and the company expects to raise again in 12 months' time.

The timing suits the state's mood. On Monday the government opened the first competitions under a £100 million sovereign AI research and development procurement scheme, run as a series of contests for AI start-ups, with one stated goal of expanding public AI compute capacity while reducing costs.

Flower Labs is not alone in chasing buyers who dislike sending data to California. Canada's Cohere has tripled its London footprint pursuing the same demand.

For a founder, the practical question is money, and it is unanswered. Licence fees for Endeavor and charges for the company's deployment tools and services have not been disclosed, so anyone weighing self-hosting against a US API bill will have to ask for a quote — and ask, too, which benchmarks produced the parity claim.