Flower Labs' Endeavor 1.1 adds image input, EU-only inference and full thinking traces

Endeavor 1.1 now accepts images as well as text[1]. Flower reports GPQA rising from 92.0 to 94.4 and HumanEval from 98.2 to 99.9 over version 1.0, and says the GPQA gain works out to about 30% fewer errors[1]. It claims the top score on its own intelligence composite, which gives equal weight to GPQA, AIME 2026, HumanEval and IFEval, ahead of leading OpenAI and Anthropic models[1]. Those are Flower's self-reported results on a mix of benchmarks it picked, so they're worth testing on your own workloads before taking at face value.

The documentation lists the model as flwrlabs/endeavor-1.1 on the managed API at api.flower.ai/v1, with support for up to a 1M-token context window, though the effective limit depends on the deployment and is confirmed during onboarding[3].

The headline for European buyers is location. Customers can now choose Flower's global service or inference on GPUs located in the EU, backed by service-level agreements[1]. Flower argues that global frontier APIs don't guarantee where a request runs, and points to Azure's UK-only per-token offering still being limited to GPT-4o and GPT-4.1 mini[1]. It plans UK-only inference next, followed by regional options in the Middle East, Asia and Africa[1].

Flower is also selling openness about the model's reasoning. Endeavor exposes readable thinking traces, and customers own every output and trace they generate and may use them to fine-tune or train their own models, including smaller specialists for recurring tasks[1]. Flower contrasts that with OpenAI showing only summaries of its thinking, Anthropic usually withholding raw traces, and both restricting the use of outputs to train competing models[1].

Endeavor 1.1 is still available by request rather than self-serve[3]. It runs either as a Flower-managed service or as a private deployment in a customer's own cloud or on premises, where the customer decides when to upgrade[2]. Model artifacts are distributed privately, with formats confirmed at onboarding, so this isn't an open-weights download[3].

Most "sovereign AI" offerings so far have meant smaller local models or older frontier models in a regional data centre. Flower is pitching a current frontier-class model that can run in the EU or behind a company's own firewall, with the reasoning visible and the outputs reusable for training. If its benchmark claims hold up under independent testing, that's a meaningful option for regulated European teams that can't send prompts to US-hosted APIs.