DOE and Arcee Are Building America's First Federal Open-Weight Science Model — Contributions Due August 25
The U.S. Department of Energy and Arcee AI have launched the Genesis Open Models Initiative and announced Genesis-Science-1 (GS1), the first open-weight foundation model built for federal scientific research under executive-order authority. The public contribution portal — hosted by Argonne National Laboratory at genesisopenmodels.anl.gov — is now accepting applications. The first track closes August 25.
Why Open Weights, and Why Now
The framing is explicit. A national laboratory running a proprietary closed model controls the hardware but not the weights. It stays dependent on that company’s pricing, roadmap, and continued cooperation. GS1 inverts that: the government backs a model whose weights it can hold, run on its own infrastructure, and preserve indefinitely without staying tethered to an external API.
Arcee CEO Mark McQuade put it directly: “A country cannot lead in AI if everything it leads in is closed.”
The initiative was authorised by executive order in November 2025 as part of DOE’s broader Genesis Mission, which aims to double the productivity and impact of American science and engineering within a decade.
The Model
GS1 will be trillion-parameter-class, built on Arcee’s next-generation Trinity model architecture. It is not a benchmark-tuned general model. The design is built around long, difficult scientific workflows: running HPC code modernisation jobs in Fortran, CUDA, and MPI; executing simulation campaigns; doing materials-science analysis; failing mid-run and recovering.
The split of labour is explicit:
- Arcee: compute, training, post-training, workbench tooling, governed execution system
- DOE national labs: scientific problem framing, reviewed training data, evaluations, results validation
GS1 operates inside sandboxed scientific workbenches with staged execution, checkpointing, and a full log of every prompt, tool call, and intermediate result. Humans approve anything touching safety, security, publication, or compute budget. The model does not get open-ended access to DOE infrastructure.
The Contribution Call
DOE opened a public portal for universities, companies, national labs, nonprofits, and research organisations to submit materials:
| Track | What’s Needed | Deadline |
|---|---|---|
| Foundation-stage data | Scientific text, code, documentation for pretraining | August 14, 2026 |
| Post-training data | SFT examples, RL tasks, held-out evals, verifiers | August 25, 2026 |
Submitted material passes a five-gate review covering scientific fit, rights clearance, and technical integration before entering the training pipeline. This is a curated pipeline, not an unrestricted upload.
What’s Not There Yet
Genesis-Science-1 is a programme, not a released model. There are no benchmark results, no architecture specification, no licence, and no firm release date beyond “later in 2026.” Weights, a technical report, and public demonstrations are promised on delivery.
For now the signal is procurement and governance: DOE is building an open-weight supply chain for scientific AI before it has proof the model works. The August 14 and 25 deadlines are the first hard data points on how much serious scientific data shows up.
What It Means for the Market
Any vendor currently selling closed AI tooling into government science programmes now faces a federal open-weight alternative that the government itself owns and shapes. Research labs that want to hold model weights locally, adapt to sensitive domains, and trace the full work trail behind a result have a new path that doesn’t depend on an API key. That changes the procurement calculus, even before GS1 ships.