Deep Principle posts Hydrogen-1.0 science model on GLM-5.3

Deep Principle, a company that builds AI for materials research, has posted the weights of Hydrogen-1.0, a large language model for scientific problem-solving, on Hugging Face. The model is built on GLM-5.3 from Z.ai, the company says on its model card. Hugging Face lists it at about 753 billion parameters.

The weights are under the MIT license, which lets anyone use, change and redistribute them. Downloading them requires agreeing to share Hugging Face account details with Deep Principle. The repository went up Saturday.

What it is for

The card says Hydrogen-1.0 is meant to help with scientific questions, explain equations and derivations, and draft answers to open-ended research problems “for expert review.” It works through text. Deep Principle says it can also support research tools built on top of it.

The company says the model was further trained on top of GLM-5.3 with supervised fine-tuning and reinforcement learning, using data that includes domain knowledge, research scenarios, laboratory records and what it calls discovery trajectories. The full datasets and training recipes are not released, the card says.

The card lists a context window of 1 million tokens and up to 128,000 tokens of output, though it says real limits depend on the server setup. Deep Principle says it tested the model on eight graphics processors using the SGLang serving software. Smaller 4-bit versions are posted separately for Nvidia’s B200 and B300 chips and AMD’s MI350X and MI355X.

What the company warns

The card is direct about limits. It says the model may produce wrong facts, calculations, equations or references, and that “a detailed or confident response does not establish correctness.” It says answers should not be used for decisions affecting health or safety. A long context does not guarantee the model uses every detail, and the 4-bit versions may answer differently from the full weights, the card says.

Examples on the company’s technical blog, including reaction-network exploration and structure analysis from nuclear magnetic resonance data, paired the model with outside tools, the card says, and “outputs require scientific review.”

The card publishes no benchmark scores. It says evaluation focused on scientific multiple-choice questions and open-ended research problems.

On its own site, Deep Principle says its mission is “to unlock breakthrough materials with AI.” It also sells Agent Mira, which it describes as an AI scientist for materials research and development.

Our read: Hydrogen-1.0 is a useful test of a growing pattern. A specialist company takes a large model from another lab and tunes it for one field. The card is more honest than most. It tells scientists not to trust confident answers and says its outputs are supporting material for users to review. What it does not offer is evidence. There are no scores on the card, no released training data and no independent test, so readers have only the company’s word that the tuning helped. Running a model this size on eight high-end chips is beyond most university labs, which limits who can check it. The MIT license is a real plus, since it lets outside researchers try. The question is whether any of them publish results.

42AI pitches a model of human behavior

42AI says on its website that it has built what it calls the first “Behavioral World Model,” which simulates and predicts how people and AI agents behave. It pitches the model to brokerages, payment firms and cybersecurity teams, and says it can predict a client’s trades up to 24 hours ahead, with response times of 1 to 10 milliseconds and 2 to 12 milliseconds for different functions.

Also today

Anthropic files an amended plan to go public. OpenAI says its Georgia campus can cut up to 1,000 megawatts of power use. Oracle launches a $20 billion stock and bond sale amid its AI buildout.

What’s next

Deep Principle says its quantized versions should be validated before use. Its technical blog has a public training overview, the card says.

Sources: Hugging Face, Deep Principle, 42AI