Endeavor 1.0 ============ Endeavor 1.0 is Flower Labs' frontier-class language model. It is designed for advanced reasoning, coding assistance, and long-horizon agent work. It is a core generalist model across products, agents, and complex workflows rather than a model narrowly optimized for a single benchmark family. Endeavor 1.0 is the second model in the Flower model program, following Lizzy 7B, the sovereign 7B model built for UK use. Endeavor 1.0 is a production-ready preview release, initially available to a select group of organizations and partners as we expand access. Here, ``production-ready`` describes the supported managed and private deployment paths; ``preview`` describes the release's limited, request-based availability. Endeavor 1.0 is available in two operating modes: - A Flower-managed service: call Endeavor 1.0 from existing applications and agent frameworks through familiar model APIs and response formats. - Private deployment: self-host Endeavor 1.0 in your own environment, with Flower support across your data, infrastructure, and systems. Private deployment is a supported operating arrangement and Endeavor 1.0 is not currently provided as a public open-weight release. At a glance ----------- .. list-table:: :header-rows: 1 * - Property - Value * - Publisher - Flower Labs * - Family - Endeavor * - Version - 1.0 * - Positioning - Frontier-class generalist model for reasoning, coding, and long-horizon agent work * - Access - Flower-managed API, or private deployment on infrastructure you control; by request during the 1.0 preview * - Model artifacts - Availability and formats are confirmed during private-deployment onboarding; no public checkpoint is listed for download during the 1.0 preview * - Predecessor - Lizzy 7B, a sovereign 7B model built for UK use Getting access -------------- During the Endeavor 1.0 preview, access is available by request as we onboard a limited number of organizations and partners. To request access, use the `Endeavor 1.0 access form `_. Preview interface and support ----------------------------- Before integrating, approved participants should confirm deployment-specific details during onboarding: the endpoint and credentials, model identifier, supported API and response formats, context limits, tool and streaming support, usage limits, and support channels. These details can vary between the Flower-managed service and a private deployment. The complete technical contract is provided only to approved participants during onboarding. It is not published as part of the public Endeavor 1.0 documentation. Architecture and configuration ------------------------------ Endeavor 1.0 is optimized as a complete AI system rather than as a set of weights alone. It is tuned for how the model reasons at inference time, how it builds and maintains context, how it uses tools, and how it checks, revises, and recovers when a step fails. These behaviors are developed and evaluated across multiple custom coding and agent harnesses, varying tool sets, interaction formats, and execution environments, so that performance transfers across real workflows rather than depending on a single benchmark setup. In the managed service, Flower operates deployment, scaling, and model operations. In a private deployment, you run Endeavor 1.0 in your own environment and choose when to adopt upgrades, retaining the option to operate Endeavor 1.0 independently of any single provider. Training approach ----------------- Endeavor 1.0 is built from a different starting point: it builds on mature, widely available capabilities from leading open-weight models, and adds the capabilities developed through the Flower model program on top of that foundation. - **Mature foundation**. General language understanding, public knowledge, and common coding patterns from leading open-weight models, so Endeavor 1.0 builds on what the wider ecosystem already does well rather than recreating it from the ground up. - **Flower-specific capabilities**. UK-specialist knowledge and reasoning carried over from Lizzy, together with new model behaviors, specialist capabilities, and training advances developed for Endeavor 1.0. - **Continual pre-training, targeted post-training, and model integration**. These stages bring the two sets of strengths together and allow Endeavor 1.0 to be refined over time. - **Signals from real enterprise work**. FlowerBench enables repeatable evaluation on real, high-value enterprise workflows without moving the underlying proprietary data. Tasks are contributed by organizations in the opt-in Flower Enterprise Evaluation Network and run inside their own environments. The resulting end-to-end signals guide Endeavor 1.0's evaluation design, harness development, post-training, and system-level improvements. Evaluation highlights --------------------- The table below places Endeavor 1.0's results on four core evaluations alongside separately reported scores for leading closed and open-weight models, including GPT-5.6 Sol, Claude Fable 5, Kimi K3, and Nemotron 3 Ultra. .. list-table:: :header-rows: 1 * - 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 Within this cross-source comparison, Endeavor 1.0's reported HumanEval score is the highest. Its reported AIME 2026 score equals those reported for GPT-5.6 Sol and Claude Fable 5; its scores exceed the reported Kimi K3 scores on three of four benchmarks (HumanEval, IFEval, AIME 2026) and the reported Nemotron 3 Ultra scores on all four. The reported GPQA scores for GPT-5.6 Sol, Claude Fable 5, and Kimi K3 exceed Endeavor 1.0's; the reported IFEval score for GPT-5.6 Sol also exceeds Endeavor 1.0's. Scores for the comparison models are as reported by the respective vendors and third-party evaluations (Artificial Analysis, Moonshot AI, NVIDIA, and Thinking Machines Lab) under different harnesses and settings; treat the cross-model comparison as indicative rather than a controlled head-to-head measurement. The Endeavor 1.0 scores reflect Flower's internal evaluation of the launch configuration available on 31 August 2026. Representative public sources used for the comparison set include the `Kimi K3 report `_, the `NVIDIA Nemotron 3 Ultra report `_, and the `Thinking Machines Lab Inkling evaluation `_. These sources do not use one common evaluation harness. Additional agentic and reasoning evaluations will be available soon. Safety and limitations ---------------------- Endeavor 1.0 can make mistakes. It can produce incorrect, outdated, incomplete, or over-confident responses. Verify important outputs before using them, and do not rely on Endeavor 1.0 as the sole basis for medical, legal, financial, safety-critical, or other high-impact decisions. Higher-risk workflows require appropriate human oversight, domain review, access controls, monitoring, and downstream moderation. Endeavor 1.0 is a broad generalist model, but its tone, assumptions, and responses can still reflect limitations or biases in its training data and system design. Use of the Flower-managed service is subject to the `Flower Acceptable Use Policy `_. Deployment-specific safety, monitoring, and support requirements will be detailed during onboarding. Next steps ---------- - To see the first model in the Flower model program, see :doc:`lizzy-7b`. - For enterprise deployments and custom work, see :doc:`enterprise`. - For product details and the datasheet, see the `Endeavor 1.0 model page `_. - For Flower Labs research, see `Flower Research `_.