# Flower Endeavor 1.1: Image Inputs, EU Inference and API Access

Flower adds vision and regional inference to its enterprise model, but access and benchmark comparisons need a closer look.

By Paras Madan | 2026-11-04

Flower released Endeavor 1.1 on November 3, adding image inputs and GPU inference in the European Union to its enterprise AI model. The release also reports stronger reasoning results. For teams choosing where their data is processed, deployment options may matter more than its headline benchmark score.

## What changed in Endeavor 1.1

The [release announcement](https://flower.ai/blog/2026-11-03-introducing-endeavor-1.1) identifies image inputs and EU inference as new capabilities. Global inference remains an option. UK-only inference and additional locations in the Middle East, Asia and Africa are plans, not current regional availability.

Flower reports GPQA Diamond increasing from 92.0 to 94.4, HumanEval from 98.2 to 99.9, and IFEval from 94.1 to 94.5. AIME 2026 stays at 99.9. These are company-reported results, not evidence that every production task improves by the same amount.

Image support expands the kind of input an enterprise workflow can accept. It does not, by itself, establish accuracy on a company's scanned documents or specialised diagrams. Those need their own tests.

## What was already available

Endeavor 1.1 builds on the [September 1 Endeavor 1.0 preview](https://flower.ai/blog/2026-09-01-introducing-endeavor-1.0). Managed and private deployment were already part of that offering, alongside control over model upgrades. They should not be mistaken for features first introduced this week.

The earlier release also described an approach combining Flower's own capabilities with open-weight foundations. That matters when interpreting the product: a privately deployable enterprise model is not necessarily a new foundation model trained entirely from scratch, or a public weight download.

For an existing evaluator, the practical question is therefore narrower than whether to adopt an entirely new platform. It is whether the additional input type, region and measured capability gains justify changing the model behind an established workflow.

## Endeavor 1.1 API access and context window

Flower's [model documentation](https://flower.ai/docs/model/endeavor.html) lists Endeavor 1.1 as available by request. The managed API uses the endpoint https://api.flower.ai/v1 and the model identifier flwrlabs/endeavor-1.1. Private deployment artifacts are arranged during onboarding; the documentation does not offer unrestricted public weights.

The documented maximum context is one million tokens, but the supported effective limit depends on the deployment. That distinction matters before designing around very large inputs. Confirm the actual configured limit, tool support, streaming behaviour and rate limits for the endpoint you receive.

The reviewed release and model documentation do not publish a per-token price. Access approval is also separate from having a working, funded production integration.

Flower provides an [OpenCode integration guide](https://flower.ai/docs/model/endeavor-opencode.html) using its API key and the OpenAI-compatible Responses interface. This gives developers a concrete starting point for evaluation. Compatibility still needs testing against the particular client features an application relies on.

## How to read Flower's benchmark claim

The headline composite gives equal weight to GPQA Diamond, AIME 2026, HumanEval and IFEval. Flower reports 97.18 for Endeavor 1.1 on that selection. This is a particular aggregation, not an all-purpose measure of intelligence.

More importantly, the documentation says comparison figures come from different sources and evaluation setups. Cross-source comparisons are indicative. A small lead in such a table cannot establish a reliable ordering for coding agents, document extraction or long-running enterprise tasks.

## What EU inference changes

EU processing can make an evaluation possible for a team whose infrastructure requirements rule out other regions. It does not automatically settle data retention, contractual protections, logging or regulatory compliance. Those remain separate questions.

The useful test is to run the same real workload against the proposed deployment and measure accuracy, failure recovery, latency and total cost. **Control over where a model runs only becomes valuable when the model also works reliably there.**

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and deployment details into stories developers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).