Arcee AI leads model development while DOE labs contribute scientific materials and evaluation. The agreement, rights and release criteria remain undisclosed.
DOE says "no federal funds are involved" in Genesis-Science-1, a trillion-parameter-class scientific model that Arcee AI plans to release publicly later in 2026. Public descriptions explain the work more clearly than the agreement behind it.
Arcee AI will secure the compute, lead model development, build the governed execution environment, and conduct evaluation and release. Participating Department of Energy laboratories will provide reviewed scientific materials, define representative research tasks, design evaluations, and validate results, according to Arcee AI's project description.
They do not identify the legal instrument governing the collaboration. DOE and Arcee AI have not publicly named the signatory laboratory or DOE entity, the ownership of the resulting weights, the rights attached to contributed materials and evaluations, or the criteria that will determine whether the model is ready for release.
The scope of DOE's funding statement is also unclear. A federal laboratory can contribute staff time, facilities, and services without transferring money to a company. DOE has not said whether "no federal funds are involved" means no funds are being paid, transferred, or awarded to Arcee AI, or whether it means no federal expenditure supports the laboratory work. The quote comes from written responses provided to Supercomputing News and has no equivalent clarification in the public project documents.
The published division of work is specific. In its public description, Arcee AI says it will curate training data, handle pretraining and post-training, create scientific workbenches from approved DOE materials, and operate the model through a governed system. Those workbenches will reproduce tool-using scientific workflows, with people retaining approval over safety, security, publication, and computing-resource decisions.
Arcee AI's description assigns participating laboratories release-reviewed scientific materials, research tasks, evaluation design, and result validation. The Argonne-hosted contribution portal distinguishes foundation-stage text, code, and documentation from post-training examples, workflow environments, held-out evaluations, tests, and verifiers. It also seeks experts to review tasks and model behavior.
Submitting an application transfers no underlying material. Applicants initially submit descriptions, metadata, expertise, proposed uses, and handling terms. Selected contributions may enter different development stages under contributor-specific terms. Calling all of that "data" would erase the portal's distinctions.
Argonne National Laboratory told SCN that the collaboration did not result from a solicitation or competitive award and that Arcee AI was the first model developer to engage. The department said talks with other developers are underway, and that companies can reach out through the Genesis Consortium, or through their call for this open science effort.
Arcee AI describes Genesis-Science-1 publicly as a sparse, trillion-parameter-class model. Chief Technology Officer Lucas Atkins told SCN that the current design targets fewer than 100 billion active parameters, which he described as a ceiling rather than a final specification. The final architecture, benchmarks, and technical report remain unpublished.
Arcee AI intends to release weights, a technical report, and public workbench and demonstration artifacts later in 2026. No date or final license is public. Open weights would let institutions hold and operate the model, but would not guarantee open training data, open-source training code, or unrestricted commercial use. Atkins told SCN that Arcee AI has no plan to restrict commercial use; that remains an interview-only statement until the license is published.
The missing agreement should allocate rights across the project. Public documents do not say who will own the base weights, derivatives, tooling, or evaluation artifacts. They also do not specify whether laboratory materials may be used for pretraining, post-training, evaluation, or some combination, or whether every selected contribution must be publicly releasable.
Genesis-Science-1 sits inside a broader DOE architecture for multiple model types, scientific agents, shared data resources, and reusable evaluation tools. It is separate from the projects named through DOE's $293 million competitive Genesis funding opportunity. That places the model in a different partnership lane, but does not show that DOE laboratories incur no personnel, facility, or internal program costs.
Arcee AI has stated an intended 2026 window, but the parties have not published the model's final specification, license, evaluation protocol, release thresholds, or treatment of failed evaluations. Those details will determine what the public receives and what researchers can do with it.
DOE laboratories can work with private organizations through technical-assistance agreements, user-facility agreements, licenses, personnel exchanges, strategic partnership projects, and cooperative research and development agreements. DOE says the appropriate mechanism depends on the objective of the partners.
A CRADA illustrates one way a laboratory can contribute substantial resources without sending funds to a private partner. Under 15 U.S.C. Section 3710a, a laboratory may provide personnel, services, facilities, equipment, intellectual property, or other resources, but it may not provide funds to the nonfederal party. The staff, facilities, and services still carry federal costs. Neither DOE nor Arcee AI has said Genesis-Science-1 operates under a CRADA.
The Genesis Mission executive order leaves several options open. It directs DOE to establish collaboration mechanisms that can include CRADAs, user-facility partnerships, or other arrangements, and it calls for standardized data-use and model-sharing agreements with policies covering ownership, licensing, trade secrets, commercialization, and data security.
Federal research organizations have combined public scientific expertise or data with private compute and model engineering before, including Lawrence Livermore and Meta on the open OPoly26 polymer dataset under a data-transfer agreement, and NASA and IBM on the Prithvi geospatial foundation model through a Space Act Agreement. Neither establishes the terms of Genesis-Science-1.
An open-weight release creates access. It does not establish that the model can complete scientific work reliably, preserve a reproducible record, or recover safely when a tool or computing service fails. Those are the capabilities Arcee AI says Genesis-Science-1 is being built to provide.
DOE is using the contribution portal to expand scientific tasks, evaluation resources, and expert-review capacity. Its gates cover scientific fit, rights and handling, evaluation readiness, technical integration, and final selection. Selected contributors are expected to work with the researchers building and evaluating the model, and the portal says they will be credited in the technical report and release materials. DOE told SCN separately that publishing a list of accepted contributors had not been decided and would require permission from partners. A credit line in a release document and a published contributor list are different commitments, and only the first has been made.
The Argonne-led EAIRA preprint describes four classes for evaluating AI research assistants: multiple-choice questions, open responses, laboratory-style experiments, and field-style experiments. It provides a framework for constructing evaluations, not a finished release gate for Genesis-Science-1.
Experience with Argonne's ChemGraph leaderboard illustrates the attribution problem. An unsuccessful workflow can reflect a model error, a tool failure, or an unavailable API. Without detailed traces, an end-to-end score cannot reliably distinguish them. Genesis-Science-1's broader tool-using design makes failure attribution, trace capture, and component-level evaluation core program requirements.
Arcee AI told SCN it will release an open-source agent harness in August under an OSI-approved license it has not yet selected. That release is an Arcee AI project rather than a joint DOE release, and it is not the Genesis-Science-1 harness. Atkins said the program's harness will be built on top of that work and released with the model, which has no announced date. Among everything the two parties have described, the August harness carries the only firm date.
The evaluation plan remains unpublished. Until DOE and Arcee AI describe the tasks, failure classes, provenance records, reviewers, and release thresholds, readers will know what the model is meant to do without knowing how success will be judged.
As of August 6, the Argonne-hosted portal lists August 14 as the application deadline for foundation-stage contributions and August 25 for post-training contributions, with deliveries due August 28 and September 14. It expects additional deadlines about every three months and invites other organizations to offer complete open-weight models as well as scientific materials and evaluation work.
Arcee AI's own project page still listed August 6 as the close of the first contribution window on the same day, in both places where it gives the date. Applicants working from the vendor page rather than the portal would have understood the window to have closed. The two pages had not been reconciled at publication.
Before the intended release later in 2026, the parties still need to identify the governing agreement, signatory, ownership and use rights, final architecture, license, and evaluation protocol. The next model partner will show whether DOE is building a repeatable program or a collection of separately governed collaborations.
The division of work is becoming clearer. The governing mechanism, rights and evaluation plan are not.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
