TL;DR

Flower Labs, spun out of Cambridge in 2023 and now split between London and Hamburg, has released a general-purpose model it calls Endeavor. The company claims comparable task performance to GPT-5.6 Sol and Claude Fable 5, and is selling on the one thing those cannot offer: the model runs on the buyer’s own hardware.

The pitch is architecture, not benchmarks

Endeavor is assembled rather than trained from nothing. Freely available models from Meta and Mistral supply much of the raw capability; Flower adds a reasoning model of its own named Lizzy, plus the harness software that turns a model into a working agent. Chief scientist Nicholas Lane, who teaches machine-learning systems at Cambridge, argues the performance gap that once justified paying for closed systems has become “vanishingly narrow”.

Confusingly, Endeavor itself is closed and licensed. What customers get is not open weights but somewhere to put them — local deployment, and federated learning that trains against sensitive records without shipping those records to anyone’s central server. That combination is aimed squarely at hospitals, banks, defence contractors and governments.

The company has raised roughly £18m ($23.6m) since passing through Y Combinator, and says its tooling is in use at over 2,500 organisations including JP Morgan, Red Hat and the US energy department.

Why this week

Lane’s framing — that Europe should not rent its intelligence indefinitely from a few American firms — is not an abstract worry here. In June, Washington and Anthropic briefly cut access to Fable 5 on national security grounds, and Downing Street had to ask for an exemption. That episode did more for the sovereign AI argument than any white paper.

The launch also lands two days after the government opened its first £100m sovereign procurement competitions and the same morning the Alan Turing Institute warned the window for such choices is narrowing. Canada’s Cohere has been chasing the same nervous buyers, tripling its London presence.

Looking forward

The claims need independent testing; matching two flagship models is asserted rather than demonstrated. More telling will be whether regulated British buyers actually pay for control they have so far only said they wanted. Flower expects to raise again in about a year, which sets a rough deadline for proving they will.