Flower adds EU inference and image inputs to Endeavor 1.1 Flower adds EU inference and image inputs to Endeavor 1.1

Regional deployment and reusable outputs

Flower says the EU option runs inference on GPUs within the European Union, alongside its global service. It also offers deployment in a customer’s private cloud or on premises. Customers can inspect readable reasoning traces and use generated material to train specialist models. UK-only inference and additional regional options remain planned expansions. Flower reports GPQA rising from 92.0 to 94.4 and HumanEval from 98.2 to 99.9 against version 1.0. Its composite equally weights four selected benchmarks.

Model limits and client settings differ

The model documentation lists a maximum context length of one million tokens. Effective request limits depend on the deployment, upstream route, output budget and internal processing overhead. Flower supplies the complete technical contract during onboarding, including the supported interfaces and usage limits. Model artifacts are privately distributed, with formats and availability settled during that process. The documentation describes development across different coding and agent harnesses, tool sets and execution environments. Its training approach combines existing open-weight capabilities with continual pretraining, targeted post-training and model integration.

A catalog connects the coding client

Flower’s Codex guide provides a model catalog that exposes Endeavor in the selector and defines its capabilities. The setup uses the Responses interface, a Flower provider entry and an API key passed through an environment variable. The CLI and desktop application share their configuration file, so the guide calls for preserving unrelated settings rather than replacing the entire file. Its example selects low reasoning effort. For local desktop tasks on macOS, the instructions launch the application from a terminal carrying the key. The guide warns that opening the application from the Dock may not supply that credential. This setup sends local coding tasks to Flower using a Flower API key.

Enterprise tasks add a different evaluation signal

Flower’s model documentation also identifies FlowerBench as an input to its development process. The company introduced that evaluation program on July 6. Participating enterprises contribute tasks involving their own tools and proprietary context. Evaluations remain inside those organizations; sanitized results are shared outside. Its insurance example combines extracting policy terms, cleaning a statement of values, running a rating model and producing quote tables and a final report. Verifiers check the required deliverables. The program measures intermediate artifacts as well as completed work, execution time, token consumption and cost. These workflow measurements address a different question from whether a model answers a public benchmark prompt correctly.