Today the Department of Energy (DOE) and Arcee AI announced the development of Genesis-Science-1 (GS1), an open model for scientific research. This is a joint effort to bring advanced AI into scientific research across a wide range of fields.

GS1 is an American open-weight model for scientific research, built together with the DOE and its national laboratories through the Genesis Mission. Arcee has secured the compute, will handle training and post-training, build the scientific workbenches and the system around the model, and prepare it for release. DOE scientists will shape which problems are worth solving, provide the data and environments the model learns from, and be the true test as to whether its work holds up under scrutiny. GS1 will be a trillion-parameter-class language model paired with a governed execution system for long, difficult scientific work, released openly later this year with the weights, a technical report, and public demonstrations. GS1 is built on top of our next generation of Trinity models.

The case for American open models

Just a year ago, Arcee made a decision that was difficult to defend. We began training our own open models from scratch, in the United States, when the faster and cheaper path pointed elsewhere. Strong open models were already available to download, and the reasonable move was to take one, adapt it, and build from there. We understood that case. It’s how we’d been operating before, after all. We went ahead anyway, because we kept seeing the need.

Some institutions can't treat a model as a service. A bank, a hospital, a university, or a national laboratory may need to keep a version stable for years, hold it to their own standards, retrain it for a narrow field, and run it on their own systems without sending sensitive data anywhere. For them, a model is more than its benchmark scores. It's the weights, the training history, the license, the certainty that it will perform reliably indefinitely, and the supply chain behind it, all the way down.

We built the Trinity models to serve institutions like these. When the Genesis Mission came along, it fit our ethos.

We have real admiration for the open-model labs in China. DeepSeek, Qwen, Kimi, MiniMax and GLM have built excellent models that people rely on, and they kept sharing open weights when much of the field was moving the other way. They earned their standing. Yet their work also showed how few capable open models were being made in the United States. For an institution handling sensitive work, capability is only part of the question. It also matters who trained a model, where, under what license, and which country's laws sit behind the company that made it. Those are fair questions, and a leaderboard doesn't answer them.

We think the right response is to build more capable open models here at home, so that institutions who need them have somewhere to turn. Closed American systems will stay valuable, and many are superb. What they can't offer a national laboratory is a model it holds in its own hands, free to preserve, adapt, and run on its own terms. We wanted American science to have that option too.

What it takes to be useful for science

Scientific computing rarely looks like a clean question with a clean answer. More often the work means aging code, a simulation that stopped halfway, logs that disagree, and a stack of results someone has to make sense of. Progress comes from a long series of small judgments about what to trust and when the evidence is finally strong enough to report.

GS1 is being built for that kind of work. It will operate inside governed scientific environments, use approved tools, write and repair code, recover when a run fails, and keep a clear record of the reasoning behind a result. People stay in charge of anything touching safety, security, publication, or how much compute is used, and the system never has open-ended access to DOE infrastructure. Open weights are what allow laboratories to govern the model directly. It can keep a known version, evaluate the model in its own environment, adapt it to its field, and follow the full chain of work behind an answer.

We are not done

Our reasons for building open models have never rested on a single project. Open models matter now more than ever, increasingly as a matter of national importance. Much of the current debate is about which foreign-built models American institutions should be allowed to use, and restrictions can help at the margins. The more durable answer is to build American models that are better than the alternatives. Since releasing Trinity Large Thinking, that’s the work we've been focused on, and GS1 is part of it. We intend to keep making the case for open models, and to keep backing it with systems people can rely on.

Help us build it

GS1 will be shaped by the people who know scientific work inside and out. The DOE is opening a contributor program for researchers, laboratories, universities, companies, and nonprofits, and we're speaking with infrastructure partners who can add training or evaluation capacity. Open source is built for everyone, so we want it informed by everyone. If your team can help move the model forward, we'd like to hear from you.

The most useful material is often messy, full of the partial runs, compiler errors, odd interfaces, and failure modes that polished benchmarks leave out, which is exactly what a serious scientific model needs to see. Contributors keep ownership of what they share and set the terms for how it's used, and the first step asks only for a description rather than sensitive files.

We kept building open models because we believed that, in time, the institutions that matter most would want to hold and shape their own AI rather than depend on someone else's. The Genesis Mission is a chance to put that to work on problems that matter, and it fits what we've been building toward better than we could have hoped. There’s still lots of work ahead, and we hope you’ll help us build in the open, starting with science.

The DOE contribution portal opens this week. Check back here for the link.