US open-weight models expand the options for self-hosted AI

Arcee AI, Poolside, Nvidia, Thinking Machines Lab and Meta offer downloadable AI models, giving US buyers more choices for deployments on infrastructure they control. Reflection AI remains a prospective entrant, with funding but no released weights, model card or published benchmark. The expanding American lineup offers alternatives for buyers with jurisdiction and licensing requirements, without establishing a clear capability-per-dollar advantage over Chinese families such as DeepSeek, Qwen and Kimi.

The available models serve different needs. Arcee’s Trinity targets general-purpose deployments under OpenMDW 1.1, following an initial Apache 2.0 release. Poolside’s Laguna focuses on private coding assistance, with licenses varying by model. Nvidia’s Nemotron offers multiple sizes and published training data, while Thinking Machines Lab’s Inkling emphasizes fine-tuning on proprietary data. Meta’s Llama retains the broadest tooling ecosystem but uses a community license with usage and naming conditions rather than a standard permissive license.

Buyers should verify the license for the exact model variant and evaluate hardware fit, operating costs and performance on their own workloads. Self-hosting addresses data-residency needs more directly than the developer’s nationality alone. Nvidia also represents a concentration risk across the American ecosystem: it supplies training hardware, publishes competing models and invests in other developers.