Companies adopting AI keep running into the same trade-off. The most capable models live behind closed APIs, where the provider decides what changes, where the computing happens and what customers may do with the results. Running an open model yourself brings control, but it also brings the burden of hosting it and understanding how it behaves. On November 3, Flower Labs announced a model meant to remove that choice. According to the company’s announcement, Endeavor 1.1 is a frontier model that customers can deploy on their own premises, inspect step by step, and use to train models of their own.
Flower Labs is best known for Flower, an open-source framework for federated learning, the practice of training AI across many separate data holders without pooling the data in one place. Endeavor is the company’s bet that the same idea of keeping data and control with its owner should apply to frontier-scale language models.
What is new in version 1.1
Flower says Endeavor 1.1 is its third model release this year, following a model called Lizzy and Endeavor 1.0 in September. The company lists the main changes as stronger reasoning for demanding agent work, image inputs alongside text, and a new option to run inference on GPUs located in the European Union.
On benchmarks, Flower reports gains over version 1.0 on GPQA, a test of graduate-level science questions (from 92.0 to 94.4), and on HumanEval, a coding test (from 98.2 to 99.9). IFEval, which measures how well a model follows instructions, edged up from 94.1 to 94.5, while the AIME 2026 maths score stayed at 99.9. The company says the GPQA gain works out to roughly 30% fewer errors. With image input now supported, Endeavor 1.1 also posts a score of 81.12 on MMMU-Pro, a multimodal reasoning benchmark.
Flower also built an “intelligence composite” that gives equal weight to GPQA, AIME 2026, HumanEval and IFEval. On that measure, the company reports Endeavor 1.1 at 97.18, ahead of GPT-5.6 Sol (96.25), Claude Fable 5 (95.30), Kimi K3 (94.83) and Nemotron 3 Ultra (92.28). These are Flower’s own figures on benchmarks it selected; several are close to their ceilings, so independent testing on real workloads will matter more than the headline ranking.
Frontier AI that stays in the region
The most concrete business change is where the model runs. Flower argues that global AI APIs do not guarantee processing inside a particular region, and it points out that Microsoft Azure’s UK-only per-token offering is still limited to older OpenAI models. Starting with Endeavor 1.1, customers can choose between Flower’s global inference service and inference on GPUs located in the EU, backed by service-level agreements.
The company says this is the start of a wider plan. It intends to add UK-only inference and regional options in the Middle East, Asia and Africa. For organisations bound by data residency rules, such as banks, insurers and public bodies, that could make a frontier model usable where it previously was not.
Owning the reasoning, not just the answer
Flower’s announcement leans hard on the idea that customers should own the work they pay for. It makes three promises:
- Readable thinking traces. Users can see the model’s step-by-step reasoning to review its approach and find where it went wrong. Flower contrasts this with OpenAI, which it says shows only summaries of its reasoning, and Anthropic, which it says typically withholds raw traces.
- Ownership of outputs and traces. Everything the model generates for a customer is the customer’s to keep and reuse.
- The right to train on them. Customers may use those outputs and traces to fine-tune or train their own models, including smaller specialist models. Flower notes that OpenAI and Anthropic restrict using their outputs to build competing models.
The company gives a practical example: an engineering team could collect verified bug fixes along with the reasoning that produced them, then train a smaller troubleshooting model tuned to its own systems. In Flower’s framing, each reviewed example becomes an asset the company keeps, rather than spending that simply disappears into a provider’s API bill.
Three ways to deploy
Endeavor 1.1 can run as a service managed by Flower, in a customer’s private cloud, or fully on premises. Flower’s argument for the last two options is about intellectual property. Prompts can reveal source code, product plans and working methods, and a provider’s promise not to train on that data does not, by itself, keep it inside the customer’s walls. Running the model in-house also lets a company keep a stable version and decide when to upgrade, rather than having a hosted model change underneath its applications. Flower says customers can start on its managed service and move to private deployment later.
What it means
For business leaders, Endeavor 1.1 is a sign that the AI market is moving beyond a simple choice between a few US-hosted APIs and do-it-yourself open models. The selling points here are less about raw intelligence than about control: where data is processed, who can see the reasoning, who owns the output and whether a company can turn its AI spending into its own models over time.
For teams in regulated industries, regional inference with service agreements could be the deciding factor. For engineering teams, access to full reasoning traces makes it easier to audit decisions and debug agent workflows. And for companies worried about dependence on a single vendor, the right to train smaller in-house models on their own outputs offers a path toward lower costs and more independence.
The caveats are real. The benchmark comparisons are self-reported, access currently runs through a request form, and Flower has not published pricing in its announcement. Organisations should test Endeavor on their own tasks, as the company itself recommends, before treating the rankings as settled.
The bottom line
Endeavor 1.1 is a pitch for frontier AI on the customer’s terms: in the region they choose, on hardware they control, with reasoning they can read and outputs they can build on. Whether it truly matches the biggest labs will be settled by independent testing, but its core argument, that users should own their AI work, is likely to shape how many enterprises judge AI vendors from here on.