The Department of Energy today announced a plan to create a new class of open-weight foundation models as part of its Genesis Mission effort to turbocharge scientific discovery and engineering productivity. The DOE unveiled the first model as part of this effort, dubbed Genesis-Science-1, but it’s looking to the AI community for more open-weight models. Interested parties better act quickly, however, as the deadlines are coming up fast.
As part of its open-weights initiative, the DOE says it’s hoping “to galvanize the scientific and AI communities around shared infrastructure” and accelerate progress in materials discovery, energy systems, earth system modeling, fusion, biology, and high-energy physics. “This effort is grounded in the principles of open science, reproducibility, and responsible AI, with the goal of lowering barriers to advanced AI capabilities for the public good,” the DOE said.
The DOE announced that it’s soliciting proposals from open-weight model creators as well as the people who would like to use the models. This includes open-weight models for use as a base model for downstream fine-tuning, as well as AI models for immediate deployment.
It would like to hear from organizations and researchers who have access to “high-quality, domain-specific scientific data” who may be interested in contributing their data to future pretraining cycles. It also would like to hear from teams that would like to use that data to fine-tune existing models for specific use cases and mission needs, such as lab assistants, simulation surrogates, and scientific copilots.
Applications for the first contribution window for pretraining efforts will close next Friday, August 14. Applications for fine-tuning efforts will close on August 25. The DOE says additional deadlines will be rolling and occur every three months. You can submit and application for the Genesis Open Models Initiative here.
Genesis-Science-1
The first model accepeted into the DOE inititive, Genesis-Science-1, was developed by Arcee Labs. In a recent blog post, Arcee CTO Lucas Atkins and CEO and Co-Founder Mark McQuade discussed what drove them to undertake the development effort and where they plan to go from here.
(FOTOGRIN/Shutterstock)
“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,” the founders wrote. “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.”
The right response is this, Atkins and McQuade wrote, is to build better open models in the U.S. That is what the San Francisco company has done with its first model, Trinity, a 400 billion parameter language model that the company released under an Apache 2.0 license.
“Closed American systems will stay valuable, and many are superb,” Atkins and McQuade wrote in the blog post. “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.
Open Vs Proprietary
The question of whether one should use open-weight models or closed-source models for AI endeavors has many variables, including capability, price, adaptability, national sovereignty, and whether one needs to run on-prem.
The top performing foundation models in the world today are largely proprietary, run in the cloud, and are not adaptable or fine-tunable to meet customers’ specific needs. OpenAI, Anthropic, and Google are widely recognized as developing the best AI models in the world with GPT, Claude, and Gemini, respectively. But they’re all closed source and come at a relatively high cost for users.
American open-weight models, meanwhile, generally don’t perform as well as the closed frontier models on most benchmark tests. While anyone can download, modify, and run models from Arcee AI, Poolside, Nvidia, Thinking Machines Lab, Meta, and Reflection AI on their own hardware, they generally don’t meet the same performance levels on benchmarks.
These are the general tradeoffs that practitioners have weighed over the first six months of 2026. But the latest open-weight models out of China buck that trend. Earlier this year, the Stanford University’s Human-Centered Artificial Intelligence (HAI) published its latest AI index that showed that Alibaba’s Qwen and DeepSeek foundation models scoring nearly as high as the proprietary American models. And in recent weeks, Moonshot AI launched Kimi K3 and Alibaba launched Qwen 3.8 tMax, which offered another jump in capability.
The emergence of these highly capable open-weight models from China spurred calls for the U.S. Government to regulate or ban these models from use in the United States. There doesn’t appear to be the political will to take such a drastic action as banning open-weight Chinese models for use by American businesses.
After all, that would signal that the Chinese have developed superior technology to the Americans, which is debatable. American officials have accused the Chinese of copying the proprietary American models through a technique known as of distillation, which is probably true. But making a technical as commonplace as distillation illegal also would seem to be a bridge too far (Prohibition didn’t last in the U.S., either).
The Risk of Chinese Models
The DOE will not tolerate Chinese models in the National Labs–for good reason, as their use does pose a security risk, even if the models are running on-prem in secure facilities.
Chinese models have been detected behaving differently when used in a government context, according CTERA CTO Aron Brand, citing a May Booz Allen study.
(Source: Shutterstock)
“Whether this behavior is deliberate or an emergent result of training data and alignment, the enterprise risk is the same: inconsistent behavior, political constraints, and the possibility of hidden backdoors or other latent behaviors,” Brand told HPCwire.
But at the same time, the government cannot cede the high ground in open-weight models to the Chinese. The good news is that the American AI community has recently made a large commitment to open-weight models.
Rise of Open-Weight Models
In July, an open letter titled “Open Weights and American AI Leadership” was signed by the CEOs of Meta, Microsoft, Nvidia, IBM, Dell, Huggingface, Palantir, Perplexity, Mistral AI, Reflection and others. Nvidia CEO Jensen Huang’s very first post on X was about the letter.
“Our AI leadership will be judged not by one frontier AI model, but by whether the United States build a strong, open ecosystem that diffuses into every sector,” the letter reads. “This is essential for creating opportunities for innovation and prosperity across the country.”
Open-weight models are superior to proprietary models in several respects, according to the “Open Weights and American AI Leadership” letter, in that they can be downloaded, inspected, modified, and run by anyone on their own hardware. They also strengthen competition by preventing AI capabilities from being used by only a handful of companies or labs, which ultimately improve capabilities and drives down costs.
The post DOE Launches Open Model Initiative for Genesis Mission appeared first on AIwire.