Flower Labs Endeavor 1.0: Frontier AI on Your Own Terms
Flower Labs launched Endeavor 1.0 today, September 1, 2026, a frontier-class generalist AI model it says matches OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5 on key benchmarks while adding a capability neither of those models offers: the option to run it entirely on infrastructure you control, with your data staying exactly where it is.
The launch is a direct response to one of the sharpest tensions in enterprise AI right now. Organizations want frontier capability. Procurement, legal, and security teams want to know exactly where the data goes. Until now, the answer to “both” has been unsatisfying: use a less capable open-weight model or accept that your sensitive data leaves the building.
Flower Labs is arguing that the answer is federated AI, and that Endeavor 1.0 is the point at which that architecture reaches frontier performance.
What Launched Today
Flower Labs is a London and Hamburg-based startup founded in 2023 with roots at Cambridge University. Its co-founder and chief scientist, Professor Nicholas Lane, holds the machine learning chair at Cambridge and has spent years building the federated learning infrastructure that now underlies Endeavor. The company has raised over £23 million in total funding and counts Y Combinator among its backers.
Endeavor 1.0 is Flower’s second model release. The first, Lizzy 7B, was a sovereign open-weight model built specifically for UK use cases. Endeavor is the step from focused sovereign model to broad generalist, competing directly with the leading closed frontier systems rather than offering a smaller, cheaper alternative.
Deployment comes in two modes:
Managed through Flower. Flower operates the model, handles scaling and model operations, and customers interact via API. This path is faster to deploy and requires no GPU infrastructure.
Self-hosted. Flower supports full deployment inside a customer’s own environment, including support across data, infrastructure, and systems. The model weights and the Flower federated infrastructure run entirely inside the customer’s perimeter.
In both cases, the federated architecture means that training and fine-tuning on customer data happens locally. Raw data never leaves its source. Only model updates are aggregated centrally. For a hospital that needs a model trained on patient records, or a bank running fraud detection on transaction data that cannot leave the institution, this distinction is the difference between a usable product and a compliance blocker.
The Benchmark Case
Flower Labs published a direct comparison against GPT-5.6 Sol, Claude Fable 5, Kimi K3, and Nemotron 3 Ultra across four standard evaluations (Flower Labs model page):
| Benchmark | Endeavor 1.0 | GPT-5.6 Sol | Claude Fable 5 | Kimi K3 | Nemotron 3 Ultra |
|---|---|---|---|---|---|
| GPQA | 92.0 | 94.1 | 92.6 | 93.5 | 86.7 |
| HumanEval | 98.2 | 95.1 | 97.0 | 96.3 | 96.3 |
| IFEval | 94.1 | 95.9 | 91.7 | 92.8 | 91.9 |
| AIME 2026 | 99.9 | 99.9 | 99.9 | 96.7 | 94.2 |
Endeavor leads on HumanEval, ties the frontier leaders on AIME 2026, trails GPT-5.6 Sol on GPQA by 2.1 points and on IFEval by 1.8 points. Against Kimi K3, it wins on three of four; against Nemotron 3 Ultra, it wins all four.
Two things are worth noting about these numbers. First, they are Flower’s own figures. Third-party independent evaluations have not yet been published. Second, Flower Labs acknowledges that Endeavor was not built around maximizing benchmark scores. The company also runs FlowerBench, its enterprise evaluation framework, which measures agents on real-world private workflows across healthcare, insurance, legal, operations, finance, and marketing tasks. Public benchmarks, it argues, miss the failure modes that matter most in production.
That second point is actually the more credible claim. Public benchmarks are saturating. Scores on AIME 2026 and HumanEval no longer distinguish between the top models in ways that predict which one a law firm or hospital will find more reliable in production. FlowerBench’s contribution is a methodology for evaluating agents on the actual task they will perform, inside the organization’s own systems, without any of the evaluation data leaving.
Why This Matters for Enterprise AI Strategy
The European and UK context for this launch is not background noise. Professor Lane told The Times: “Europe should not have to rent its intelligence indefinitely from a handful of US companies.” That framing is not just marketing. It describes a real procurement reality that procurement offices across the EU, UK, Gulf, and Asia Pacific are now navigating.
For enterprise AI leaders, Flower Labs Endeavor raises three concrete questions.
First: does your current AI strategy depend on data leaving the building? Most closed API deployments require sending prompts and context to a US-hosted endpoint. For general business intelligence tasks, that may be acceptable. For tasks touching patient data, legal documents, financial records, or intellectual property, it often is not, or increasingly will not be as data regulations tighten globally. Federated AI eliminates that exposure at the architecture level rather than through contractual assurances.
Second: can you train on your proprietary data? This is the deeper competitive advantage Flower Labs is selling. Closing the loop from a general frontier model to one that has actually learned from your internal data, without transferring that data, has been technically difficult. Flower has spent three years building the distributed training infrastructure that makes it feasible. Clients including NHS and JP Morgan are already using it for privacy-preserving fraud detection and clinical research (Red Hat partnership blog, March 2026).
Third: what is your plan for AI that keeps improving without adding vendor dependency? Endeavor 1.0’s pitch is not a static model endpoint. It is explicitly a foundation for building “agents, evaluations, data pipelines, and improvement loops” around the model over time. The goal is for the model to become increasingly differentiated around each customer’s work, not a commodity API call.
This is the same strategic framing that Thomson Reuters used when it justified its $40 million investment to train the Thomson 1.0 model on its own legal content: the organization’s proprietary data is the moat, not the model itself. Flower is offering a path to that outcome for organizations that cannot afford a training run of their own, by using federated learning to embed the organization’s data advantage into a shared frontier base.
Infrastructure and Ecosystem
Flower’s federated learning framework is already production-grade. It runs on Red Hat OpenShift and integrates with Open Cluster Management for multi-cluster Kubernetes deployments, which matters for enterprises with data distributed across multiple cloud regions or air-gapped facilities. Samsung and Nokia use it for on-device machine learning. Banking Circle uses it for federated fraud detection. The Red Hat integration means that organizations already running OpenShift can deploy Flower’s federated infrastructure on their existing clusters without additional network topology changes.
Endeavor 1.0 is available under a commercial licence. Clients can train the model on their own data without transferring it to Flower’s systems, and the licence allows fine-tuning. This is distinct from both the fully closed model of the frontier US labs and the fully open Apache 2.0 approach of models like Tencent Hy4 or Alibaba Qwen3.8-27B. Flower’s position is in the middle: commercially licensed, deployable on your infrastructure, with federated fine-tuning support.
The launch is currently limited to a select group of organisations as Flower ramps up its compute capacity. A broader rollout is planned.
Who Should Be Paying Attention
If your organization is in a regulated vertical such as healthcare, financial services, legal, government, or defence, Flower Labs Endeavor 1.0 belongs on your AI model evaluation shortlist. The federated architecture addresses the compliance constraints that block most frontier model deployments in these sectors.
If you are in an enterprise that has been frustrated with the gap between a model’s benchmark performance and its performance on your actual workflows, FlowerBench is worth examining as an evaluation methodology. It is the only published enterprise benchmark that operates on private data inside member organizations, rather than on publicly available test sets.
If you are currently evaluating AI sovereignty strategies and weighing sovereign model deployment against managed closed API access, Endeavor 1.0 represents a new third option: frontier performance, federated architecture, commercial licence, deployed on your own infrastructure.
The gap between “AI capability” and “AI I can actually use with my most sensitive data” has been one of the primary blockers to enterprise AI adoption at scale. Flower Labs is making a credible technical argument that federated learning narrows that gap to near zero.
Whether independent benchmarks will validate the performance claims, and whether the model can hold up in the production environments of regulated enterprises, remains to be seen. But the launch marks the point at which sovereign AI stopped being a second-tier compromise and started making a case on raw capability.
If you want to assess how Endeavor 1.0 and federated AI architectures fit into your organization’s AI strategy, we work through exactly these tradeoffs with enterprise teams.