DOE and Arcee partner on open-weight model for scientific research

Why it matters

Arcee says it will provide compute and manage training, while DOE scientists and national laboratories will define research problems, provide data and research environments, and evaluate the model. GS1's value as a public scientific asset will depend on the license, provenance disclosures and reproducibility terms accompanying its release.

DOE and Arcee announced Genesis-Science-1 on July 22, 2026, with Arcee AI co-founder and CEO Mark McQuade involved in developing DOE's first open-weight model built specifically for scientific research. The contribution portal is now soliciting data, software, research environments and evaluations for the planned model. The launch announcement listed August 14, 2026, as the first foundation-stage contribution deadline, although a live version of the program homepage displayed August 6. The separate post-training application window remains open through August 25.

The Genesis Open Models initiative gives McQuade and Arcee AI a role building core AI infrastructure for the federal government's national laboratories. Arcee is the first industry partner in the program, according to DOE's July 22 announcement.

For McQuade, Genesis-Science-1 extends the bet behind Arcee, which he co-founded in 2023 with Brian Benedict and Jacob Solawetz. McQuade came from Hugging Face, while Solawetz previously worked at Roboflow. Arcee's public founding thesis centered on models that enterprises could own, inspect and run inside their security perimeter. Arcee initially focused on specialized models for private infrastructure, then moved into training and releasing its own open-weight foundation models.

Arcee's earlier financing included a $5.5 million seed round from WndrCo, Long Journey Ventures and Flybridge, followed by a $24 million Series A led by Emergence Capital. Those rounds put its disclosed funding at a minimum of $29.5 million before a later strategic funding round led by Prosperity7 Ventures and Microsoft's M12. Hitachi Ventures, Wipro, JC2 Ventures, Samsung Next and Albert Sebag also participated; Arcee did not publish the round's size.

That model-control thesis matters here. DOE laboratories need to preserve model versions, run them on government-controlled infrastructure and adapt them without permanent reliance on an outside API. The partnership will test whether Arcee can apply that approach to scientific workflows that require traceable data, repeatable execution and review by domain experts.

A startup thesis becomes federal infrastructure

Arcee says Genesis-Science-1 will be a trillion-parameter-class language model released later in 2026 with its weights, a technical report and public demonstrations. Arcee says it has secured the compute and will manage training, post-training, scientific workbenches and release preparation. DOE scientists and national laboratories are expected to define research problems, provide data and research environments, and judge whether the model's work survives scientific scrutiny.

The division of labor gives Arcee scientific problems, data, research environments and evaluations defined by DOE scientists and national laboratories. DOE gets a private partner that has already trained and released open-weight systems while preserving the option to operate the resulting model inside government infrastructure.

The federal partnership tests whether Arcee can coordinate with subject-matter experts, data stewards and infrastructure operators at national-laboratory scale. Adding parameters alone will not solve an inherited Fortran codebase, a failed simulation or a disputed experimental result.

DOE wants the messy parts of scientific work

The contribution program is designed around those harder cases. DOE is soliciting scientific text, code, documentation and structured collections for pretraining, midtraining and context extension. The post-training track covers expert demonstrations, annotated tasks, research software, reinforcement-learning environments, held-out evaluations, scoring rubrics, tests and verifiers.

DOE's launch notice listed August 14, 2026, as the first foundation-stage application deadline, with selected materials due August 28. The live program homepage subsequently displayed August 6 for that application round, creating a discrepancy in the public schedule. Post-training applications are due August 25, followed by a September 14 delivery deadline. DOE says the initial application collects descriptions and metadata rather than scientific files. Contributors specify proposed usage terms before material changes hands.

DOE says applications will pass through reviews covering scientific fit, rights and handling, expert readiness, technical integration and final program selection. The Genesis Open Models application page provides the contribution details.

DOE's GS1 description covers high-performance-computing code modernization, experimental analysis, simulation campaigns, materials science and energy systems. Environments may include Python, Fortran, C and C++, MPI, OpenMP, CUDA, HIP, notebooks, simulation packages and computing schedulers.

Arcee describes workbenches designed to plan workflows, use approved tools, preserve task state, recover from failures and record code changes, datasets, intermediate artifacts and conclusions. Human reviewers will retain control over safety, security, publication and resource-use decisions.

Genesis enters a crowded AI-for-science field

Periodic Labs offers one comparison. Periodic Labs reportedly raised a $300 million seed round to pair AI models with robotic laboratories and experimental feedback loops, according to TechCrunch. Genesis-Science-1 is aimed at a different layer: a general scientific model and governed execution system developed with DOE researchers and intended for release with open weights.

The institutional distinction is substantial. The Genesis Mission links 17 national laboratories with scientific data, advanced computing and research facilities. That gives GS1 a government-backed route to domain-specific problems, expert evaluation and deployment on national-lab infrastructure. Periodic's strategy centers on building its own closed-loop experimental facilities with venture financing.

Other projects set a higher bar for openness. Oumi describes its model-development platform as unconditionally open across code, data, weights and collaboration infrastructure. Latent Labs concentrates on generative biology and protein design rather than the cross-domain scientific workflows DOE has outlined for GS1.

Open weights are the start of the test

Open weights give laboratories a route to inspect, preserve, run and adapt the model. Reproducibility will depend on the license, training-data provenance and release documentation that accompany it.

Those release terms will determine how broadly researchers outside the national-lab system can reproduce and extend the work. Arcee has promised weights, a technical report and public demonstrations. Access to training code, detailed data documentation and permissive reuse rights would make GS1 a substantially stronger public asset.

The open-model program is one piece of the broader Genesis Mission, which was established by executive order on November 24, 2025. DOE says the mission connects 17 national laboratories and aims to double the productivity and impact of American science and engineering within a decade.

In a July 22 announcement, DOE said Genesis Mission partners had committed more than $800 million in support. DOE said those commitments included compute resources and credits, AI models, cloud infrastructure, scientific expertise, research partnerships and direct funding.

For Arcee, the immediate task is execution. McQuade's founding argument was that important institutions would eventually demand models they could hold, modify and govern themselves. DOE has handed Arcee a national-scale venue to prove that argument on scientific work where an impressive answer is insufficient unless another researcher can inspect how it was produced.