Endeavor 1.1: Flower Labs' Frontier AI Model Now Runs on EU Servers

Most frontier AI models run where their American vendors put them. Flower Labs, a Cambridge University spinout, thinks that is a problem, and its chief scientist told The Times that Europe should not have to rent its intelligence indefinitely from a handful of US companies, as reported by Knight Frank in September. On November 3, 2026, the company released Endeavor 1.1, the third version of its frontier-class model this year, and put a concrete option behind the argument: inference on GPUs located inside the European Union.

Version 1.1 also adds image inputs, readable reasoning traces you can keep, and the right to use those traces to train your own models. Flower reports its best benchmark scores yet, though every number comes from its own testing. There is no published price: access is by request, with enterprise SLAs. Here is what the release changes, what the numbers actually say, and who it is for.

What Endeavor 1.1 actually is

Endeavor 1.1 is a frontier-class generalist model for reasoning, coding and long-horizon agent work, according to Flower's November 3 announcement. It is the company's third model release of 2026, following Lizzy (the smaller sovereign model Flower released earlier in the year) and Endeavor 1.0, which launched on September 1, 2026.

One clarification matters up front: Endeavor is not open weights. It ships as a licensed, closed model. What customers get instead of weights is deployment choice: Flower-managed service, private cloud, or on-premises installation, plus the reasoning traces and output ownership that closed APIs typically withhold.

What's new in 1.1

• EU inference. Customers can now choose inference on GPUs located specifically in the European Union, alongside Flower's global offering. The company promises service-level agreements and deployment support on the EU route, and says UK-only inference plus regional options for the Middle East, Asia and Africa will follow.
• Image inputs. The model now accepts images alongside text, making it multimodal for the first time.
• Stronger reasoning and coding. Flower reports GPQA rising from 92.0 to 94.4, HumanEval from 98.2 to 99.9, and IFEval from 94.1 to 94.5 against version 1.0, with AIME 2026 holding at 99.9. The company says the GPQA move alone means about 30% fewer errors.
• One-million-token context. The model documentation lists a maximum context length of one million tokens, per S3DAT's technical read of the release.
• Readable thinking traces. Customers can inspect the model's reasoning to review its approach, investigate decisions and find where it went wrong.
• Ownership of outputs. You keep the outputs and traces you generate. Flower explicitly permits using them to fine-tune or train your own models, including smaller specialist models for recurring tasks.
• Coding-client setup. Flower's Codex guide exposes Endeavor in the coding client's model selector, configured through the Responses interface with an API key in an environment variable.

The benchmark story, read carefully

Flower's headline claim is a composite "intelligence" score: Endeavor 1.1 at 97.18, ahead of GPT-5.6 Sol at 96.25, Claude Fable 5 at 95.30, Kimi K3 at 94.83 and Nemotron 3 Ultra at 92.28. The composite gives equal weight to exactly four benchmarks: GPQA, AIME 2026, HumanEval and IFEval, covering science, mathematics, coding and instruction following. An MMMU-Pro score of 81.12 is reported separately and is not part of the composite.

Two things to keep in view. First, this is Flower's own composite, built from Flower's own test runs. A four-benchmark average is a narrow lens, and the company chose which benchmarks to include. Second, no independent benchmark organization has published a reproduction of these results. The table below uses Flower's figures with that caveat attached.

The real product is control, not the leaderboard

Flower's launch post makes its sharpest arguments about usage terms rather than scores. Closed frontier APIs, it says, leave the model company in charge of availability, usage terms and changes to the model your AI depends on. It notes that OpenAI restricts using outputs to train competing models, that OpenAI exposes only summaries of its thinking traces, and that Anthropic typically withholds raw thinking traces. Those are Flower's characterizations of its rivals' terms, included here as the company's stated motivation rather than independent findings.

Endeavor's answer is threefold: deploy the model yourself so sensitive work never leaves your environment, inspect readable traces when something goes wrong, and keep the outputs to build on. The federated-learning heritage shows. Flower built its name on training models against data that stays where it is rather than moving to a central database, as Dealroom reported at the 1.0 launch.

The company also runs FlowerBench, an evaluation program introduced on July 6 in which participating enterprises contribute tasks involving their own tools and proprietary context. Evaluations stay inside those organizations; only sanitized results are shared. That is a different, and arguably more useful, signal than public benchmark prompts, though the sanitized results are still Flower's to publish.

What it costs

There is no published price for Endeavor 1.1, and that is worth stating plainly rather than skipping. Access is by request, with enterprise SLAs and deployment support from Flower. The real budget has two parts: whatever enterprise contract you negotiate, and the infrastructure if you choose private cloud or on-premises deployment. Flower's pitch is that this replaces per-token bills to a US lab with a controllable deployment. Until the company publishes rates, buyers cannot compare the two directly. That opacity is the main pricing catch in this release.

The company behind it

Flower Labs is a Cambridge University spinout based in London and Hamburg, founded in 2023. It has raised over £23 million (about $31 million at November 4, 2026 market rates), per Tech.eu's September reporting, and names the NHS and JPMorgan among its clients, organizations whose data-protection and data-locality rules make cloud-only AI hard to justify internally. Chief scientist Nicholas Lane, a Cambridge machine-learning professor, has framed the mission in explicitly political terms: Europe should not rent its intelligence indefinitely from US vendors.

Rather than training a frontier model entirely from scratch, Flower builds on capabilities from the open-weight ecosystem, then adds continual pre-training, targeted post-training and its own system integration: the approach it describes as assembling rather than inventing the raw capability.

Where Endeavor fits

For readers tracking the sovereignty lane, Endeavor sits between the Swiss model play (Giotto) and the sovereign-cloud play (Polarise): a frontier model you can run inside your own boundary, rather than a national model or a national cloud.

Who should care — and who should skip

This release is aimed at banks, hospitals, government departments and regulated enterprises, the buyers whose procurement teams ask where the data goes before anything else. If your workload involves patient records, transaction histories or classified material under data-locality rules, the combination of EU inference and on-premises deployment is the actual news here, not the composite score.

If you are an independent developer looking for a cheap API to build on, this is not your product. There is no self-serve tier, no published per-token rate, and no free trial mentioned in the launch material. Flower is selling control to institutions, not tokens to individuals.

Catches and open questions

• No independent verification. Every benchmark figure is Flower-reported. Treat the 97.18 composite as the company's best case until someone else reproduces it.
• EU inference is brand new. Regional deployment launched with this version; it has no track record at production scale yet.
• Closed and licensed. You get deployment choice, not weights. Do not confuse this with an open-weights release.
• Rival characterizations are Flower's. The claims about Azure's UK-only offering and about OpenAI/Anthropic trace policies come from Flower's launch post: useful context for its strategy, not verified comparisons.
• Fast release cadence. Three models in one year means production users should plan for churn; Flower's answer is that on-premises deployments let you choose when to upgrade.
• Modest 1.0-to-1.1 gains on some measures. IFEval moved only 94.1 to 94.5 and AIME 2026 did not move at all. The step forward is real on GPQA and HumanEval, incremental elsewhere.

FAQ

Is Endeavor 1.1 open source?

No. It is a licensed, closed model. Flower's offer is deployment flexibility (managed service, private cloud or on-premises) rather than downloadable weights.

How much does Endeavor 1.1 cost?

Flower has published no price. Access is by request, with enterprise SLAs and deployment support. Expect an enterprise contract; the private-deployment infrastructure bill sits on top.

What does EU inference mean in practice?

Inference runs on GPUs physically located in the European Union, alongside Flower's global offering. For organizations with data-residency requirements, the model and the workload can stay inside the EU under Flower's SLAs.

Can I train my own models on Endeavor's outputs?

Yes, explicitly. Flower permits retaining inference outputs and reasoning traces and using them to fine-tune or train your own models, including smaller specialists for recurring tasks. That is a deliberate contrast with vendors that restrict training competing models on their outputs.

How does Endeavor 1.1 compare to GPT-5.6 Sol?

On Flower's own four-benchmark composite, Endeavor 1.1 scores 97.18 against Sol's 96.25. Both numbers are company-reported with no independent reproduction, so treat the comparison as Flower's claim rather than an established result.

Sources

• Flower Labs official announcement, "Introducing Endeavor 1.1" — https://flower.ai/blog/2026-11-03-introducing-endeavor-1.1 (November 3, 2026; launch details, benchmark figures, deployment options; verified live November 4, 2026)
• S3DAT News, "Flower adds EU inference and image inputs to Endeavor 1.1" — https://news.s3dat.com/stories/flower-endeavor-1-1-eu-image-inputs (November 3, 2026; 1M context, Codex guide, FlowerBench; verified live November 4, 2026)
• Tech.eu, "Cambridge University spinout launches AI model 'competitive' with OpenAI and Anthropic" — https://tech.eu/2026/09/01/cambridge-university-spinout-launches-ai-model-competitive-with-openai-and-anthropic/ (September 1, 2026; £23M raise, NHS/JP Morgan clients)
• Dealroom, "Cambridge spinout Flower Labs launches AI model to rival OpenAI, after £23M raised" — https://dealroom.co/news/148092-cambridge-spinout-flower-labs-launches-ai-model-to-rival-openai-after-23/ (September 1, 2026; federated-learning deployment model)
• Knight Frank, "The New Frontier" weekly update — https://www.knightfrank.co.uk/research/article/2026/9/the-new-frontier---your-weekly-science-and-innovation-update---4th-september-2026 (September 2026; Nicholas Lane "rent its intelligence" quote via The Times; Lizzy background)

This is general information, not procurement or legal advice.