The U.S. Department of Energy announced on August 8 the launch of the Genesis Open Models Initiative, partnering with Arcee AI on its first model, Genesis-Science-1 (GS1), a trillion-parameter-class open-weight model built for scientific research. A contribution portal opened the same day, with the first application window closing August 14.
What GS1 is
GS1 is the first open-weight model under DOE's Genesis Mission program. According to Arcee AI's announcement, it is a trillion-parameter-class language model paired with a governed execution system for long, difficult scientific work. It is built on Arcee's next-generation Trinity model family, and the weights, a technical report, and public demonstrations are all planned for release later this year.
GS1 is not positioned as an ordinary chatbot. It targets scientific workflows: operating inside governed scientific environments, using approved tools, writing and repairing code, recovering when a run fails, and keeping a clear record of the reasoning behind a result. Humans stay in charge of anything touching safety, security, publication, and compute usage, and the system never gets open-ended access to DOE infrastructure.
Which research fields it targets
DOE lists materials discovery, energy systems, earth systems modeling, fusion, biology, and high-energy physics as target domains. Laboratories can hold and govern the model directly: keeping a known version, evaluating it in their own environment, adapting it to a specific field, and following the full chain of work behind an answer. That is what open weights enable.
Why the U.S. government is stepping in
Arcee AI is direct about the motivation: capable American open-weight models are scarce. With the Llama series abandoned, the field mostly runs on Gemma and GPT-OSS. The announcement acknowledges that Chinese open-model labs like DeepSeek, Qwen, Kimi, MiniMax, and GLM have built excellent models people rely on and kept sharing weights, and that they earned their standing. But for institutions handling sensitive work, capability is only part of the question: who trained a model, in which country, under what license, and whose laws sit behind the company that made it. A leaderboard does not answer those questions.
Arcee began training its own open models from scratch in the United States a year ago, a decision that was hard to defend at the time. GS1 continues that path. Arcee secured the compute and handles training, post-training, scientific workbenches, and release preparation; DOE scientists decide which problems are worth solving, provide data and environments, and serve as the real test of whether the model's work holds up under scrutiny.
How to participate
DOE opened the contribution portal (genesisopenmodels.anl.gov) today, inviting universities, national laboratories, companies, scientific nonprofits, and research organizations to contribute data, research environments, evaluations, and technical expertise. The first application window closes August 14. Contributions of interest include open-weight base models, high-quality domain-specific scientific data, fine-tuning and domain-adaptation teams, and evaluation and verification work.