Flower Labs Launches Endeavor as Intelligence You Can Keep

On 1 September 2026, Flower Labs released Endeavor 1.0, a generalist model it says can run on a customer’s own systems and still match OpenAI’s GPT-5.6 Sol on AIME 2026. The Cambridge spinout is not asking buyers to pick a new chatbot API so much as to stop treating frontier intelligence as a monthly rental.

Access is a preview for a short list of organisations while the company adds compute. The scores are Flower’s. The weights are not on a public download page.

Flower Puts a Frontier Model Inside the Customer’s Walls

Flower Labs, based in London and Hamburg, published Endeavor 1.0 as a production-ready preview with two ways to run it. A Flower-operated service is the fast path. A private install keeps the model and the workload on machines the customer already controls. Firms can mix the two, putting everyday jobs on Flower’s service and holding the sensitive ones in-house, then moving more of the stack later.

The company was formed in March 2023 around Flower, the open-source software for training models on data that never leaves its original site. Chief executive Daniel J. Beutel, chief operating officer Taner Topal, and chief scientist Professor Nicholas Lane, who also holds a machine-learning chair at Cambridge, built that project as researchers from 2020. The commercial entity is a Y Combinator Winter 2023 graduate, later backed by a $3.6 million seed on 8 August 2023 and a $20 million Series A in February 2024 led by Felicis, a round that valued the firm at $100 million. The two rounds total $23.6 million.

Lane put the political case in a single line.

Europe should not have to rent its intelligence indefinitely from a handful of US companies.

Nicholas Lane, co-founder and chief scientist, Flower Labs

On the same morning, Flower’s account posted the launch clip and the four headline scores.

https://x.com/flwrlabs/status/2094692986919817250

The blog is explicit about the commercial goal. Flower wants customers to start with frontier capability and build agents, tests, data pipes, and improvement loops around the model, then keep that work if they later take the install private. The model is the wedge. The durable product is the stack that grows around it.

What Endeavor 1.0 Scores Against GPT-5.6 Sol

Flower reports four core tests on the launch configuration it evaluated on 31 August 2026. Endeavor 1.0 posts 98.2 on HumanEval, the highest figure in that comparison, and 99.9 on AIME 2026, matching GPT-5.6 Sol and Anthropic’s Claude Fable 5. It trails GPT-5.6 Sol on GPQA (92.0 against 94.1) and on IFEval (94.1 against 95.9). It beats Moonshot’s Kimi K3 on three of the four tests and NVIDIA’s Nemotron 3 Ultra on all four.

THE LAUNCH COMPARISON FLOWER PUBLISHED

Those rival figures are taken from vendor write-ups and third-party tests, including Moonshot, NVIDIA, Artificial Analysis, and Thinking Machines Lab, each run under its own harness. Flower’s own docs tell readers to treat the grid as indicative rather than a controlled head-to-head. That is the right reading. The table shows a model in the same band as the US closed systems on maths and coding, not a bake-off in one lab. Lane wrote on 1 September that early preview users sat through informal blind tests against OpenAI and Anthropic models and, from what he had seen, “they can’t tell the difference.” That is a founder’s impression, not a published protocol.

Hospitals and Banks Cannot Export the Training Data

The buyers Flower already knows are the ones that cannot pour records into a US API and still satisfy a regulator. The firm’s federated software, which trains a model by sending updates rather than raw files, is used by the NHS and J.P. Morgan. LinkedIn copy for the company also lists the US Department of Energy, Owkin, and Red Hat. None of those names has been published as an Endeavor 1.0 customer. They are the reason a local generalist is even a product.

WHO NEEDS THE MODEL TO STAY PUT

• Health systems: Patient files have to remain inside hospital and NHS systems, which is why Flower has long sold training that never centralises the records.
• Banks: J.P. Morgan is named as a Flower user, the kind of shop that can fine-tune on internal ledgers only if the data never crosses a vendor boundary.
• Public bodies: Government and energy users show up in Flower’s own client list, where procurement rules and classification sit above convenience.

Anil Madhavapeddy, a professor of planetary computing at Cambridge, replied to the launch by saying he wanted it on the lab GPUs. That is the demand signal in miniature. A departmental cluster is exactly the sort of place a “frontier” model has to live if the alternative is an overseas endpoint.

On 11 September, Flower posted a walkthrough in which Endeavor inspects a broken federated job, three local clients training a blood-cell classifier, with two planted faults (too many clients required, and a learning rate set too high). The model repairs the config and checks the fix with a fresh run. That is the specimen. The pitch is a generalist that can debug the very training setup Flower sells, without the traces leaving the site.

The 7B Model That Came First

Endeavor is the second model in Flower’s programme. Lizzy 7B, an open-weight model built in the UK for UK institutions, language, and public-sector work, was posted on 15 April 2026. Flower describes the gap as four months. Lizzy is the small, downloadable proof of local control. Endeavor is the attempt to take that idea up to the closed US labs.

FROM OPEN-SOURCE TOOL TO GATED GENERALIST

1. 2020: Beutel, Topal, and Lane release Flower as research software for training on data that stays in place.
2. March 2023: Flower Labs is formed to commercialise the project after Y Combinator.
3. 8 August 2023: The company raises $3.6 million in seed funding.
4. February 2024: Felicis leads a $20 million Series A at a $100 million valuation.
5. 15 April 2026: Lizzy 7B goes out as an open-weight UK model, with the Lizzy 7B checkpoint on Hugging Face.
6. 1 September 2026: Endeavor 1.0 opens as a preview, managed or private, with no public checkpoint.

Lizzy’s own table, published by Flower, put it ahead of EuroLLM 9B and Apertus 8B on MATH (77.9 against 31.3 and 22.4) and on GPQA (34.6 against 26.8 and 28.1). Those are 7B-class scores against other European open models, not against GPT-5.6 Sol. The jump Flower is claiming with Endeavor is a different league, and a different distribution model. Lizzy you can download. Endeavor you request.

Endeavor Builds on Open Weights, Then Adds Its Own Harness

Flower does not pretend it trained a frontier model from a blank corpus. The launch post says Endeavor draws on mature open-weight models for language, public knowledge, and ordinary coding patterns, then adds UK-specialist behaviour carried over from Lizzy plus new training, post-training, and integration work. Continual pre-training is how those pieces are glued together. The company also argues that the weights are only part of the system: how the model spends reasoning effort, holds context, calls tools, and recovers when a step fails is tuned across several coding and agent harnesses rather than one leaderboard script.

A second signal, FlowerBench, runs tasks contributed by firms in an opt-in enterprise network, inside those firms’ own environments, so proprietary files never move. Sanitised outcomes then feed post-training. That is federated evaluation dressed as product development. It also explains why a company that spent years on data-that-stays-put thinks it can compete with labs that vacuum the public web. The scarce input is no longer Common Crawl. It is the hospital note, the ledger, the internal ticket, sitting behind a firewall.

The honest limit is the same fact restated as a method. If the base capability is borrowed from open-weight models that were themselves trained in large US and Chinese clusters, “sovereign” describes where the finished system runs and who may improve it, not where every floating-point operation in pre-training took place.

Why Flower Still Withholds the Weights

Lizzy shipped files. Endeavor does not. Flower’s docs say private deployment is a supported arrangement and that Endeavor 1.0 is not a public open-weight release. Artifact formats are confirmed during onboarding. The public pages list no parameter count, no training-compute figure, and no named base checkpoint. Preview users must confirm endpoints, context limits, tool support, and usage caps with Flower, because those details differ between the managed service and a private install and are not in the public contract.

WHAT WE KNOW

• Two run modes: A Flower-managed API and a self-hosted install are both on offer during the preview, by request.
• In-house scores: The four published numbers come from Flower’s 31 August 2026 launch configuration, not from a shared third-party harness.
• No public file: There is no listed downloadable checkpoint for Endeavor 1.0, unlike Lizzy 7B.

WHAT IS UNCONFIRMED

• Scale: Parameter count, training tokens, and cluster size have not been published.
• Base model: Flower has not named the open-weight checkpoint it built on.
• Named buyers: NHS and J.P. Morgan are Flower-platform users; they have not been announced as Endeavor 1.0 accounts.

The company is frank about the bottleneck. It is onboarding a limited set of organisations “as we also expand compute availability,” and on 4 September it told a requester it was “working on enabling it to as many users as possible asap.” A model you can theoretically run locally is still gated by how many GPUs Flower can stand up, and by how many private installs it can support at once. Ownership is the offer. Capacity is the constraint.

Britain’s £500 Million Bet on Homegrown Models

Lane’s rent line lands in a funding climate Whitehall has already tried to build. The government’s AI Opportunities Action Plan one-year update describes a Sovereign AI Unit backed by up to £500 million of government funding for UK AI firms, plus a commitment of £2 billion to expand public compute twentyfold by 2030, with Isambard-AI switched on in July 2025. In June 2026 it added a £1.1 billion AI Hardware Plan, including £750 million toward a national supercomputer. Flower has not been named as a recipient of that unit’s equity. The policy still explains why a Hamburg-and-London lab with a Cambridge professor on the masthead is pitching “frontier AI, under your control” now, rather than another thin open-weight 7B.

The second-order bet is therefore narrower than the press-release rivalry with OpenAI. If a bank or a trust can start on Flower’s service, wrap agents and evals around Endeavor, and later lift the same model onto its own GPUs, the switching cost flips. The intelligence becomes an internal asset. That only holds if the preview scores survive contact with those internal tasks, if private installs are actually delivered, and if Flower can find the compute it is still ramping. On 11 September the public proof was still a demo of a repaired blood-cell training run, and a form for requesting access.