Lucas Atkins is chief technology officer and head of research at Arcee AI, where he develops open-weight foundation models that organizations can inspect, customize, and operate independently. His work extends from model compression and large-scale pretraining to the agent infrastructure needed to turn those models into durable products.
Earlier in his career, Atkins developed automotive text-to-speech systems for manufacturers including Volkswagen, BMW, Ford, and Hyundai, worked on specialized translation systems for the United Arab Emirates, and collaborated with AMD on optimizing enterprise GPUs for model training.
At Arcee, he initially focused on smaller, specialized models. In 2024, he helped release open training datasets and introduced DistillKit, a framework that transfers capabilities from larger models into smaller systems using teachers’ output probabilities and internal representations. He subsequently co-launched Arcee’s hosted model engine and, with Charles Goddard, expanded MergeKit’s model-merging capabilities. By April 2025, he was co-leading Arcee Labs.
Arcee then shifted from adapting existing models to training its own. Atkins positioned the Trinity model family around control over training, deployment, licensing, and model provenance. Its flagship, Trinity Large, is a sparse mixture-of-experts model with 400 billion total parameters and roughly 13 billion activated per token. Atkins co-authored its technical report, which describes a 17-trillion-token training run, attention architecture, Muon optimization, and expert balancing. Trinity-Large-Thinking extended that foundation toward multi-turn tool use and sustained agentic work.
- Model ownership as operational trust. Atkins separates inspectability and dependable access from claims that open models are automatically safe. Controlling model weights and deployment helps organizations understand what they are running, anticipate costs, and avoid losing access when providers change products or pricing.
- Models shaped around their products. He argues that effective applications should adapt models to their specific harnesses and workflows. Open systems also let developers retain production traces and use them to improve subsequent models.
- Sparse scale with practical economics. Trinity’s architecture activates only a fraction of its total parameters per token, demanding careful routing and expert balancing while making large-model capability more practical to deploy.
- Persistent state for long-running agents. In August 2026, Atkins co-introduced NAC, an open-source runtime that coordinates parallel workers through durable workstreams and structured episodes, preserving continuity across extended engineering tasks.
Atkins also sees proprietary and open models coexisting: hosted systems remain useful, while independently operated models give builders a lasting alternative when control, customization, and predictable costs matter.
