Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery

Much of the recent AI-for-science progress has happened behind closed APIs. Deep Principle, a Hangzhou-based AI-for-materials company, is releasing Hydrogen-1.0 with open weights under the MIT License, aimed at chemistry, materials science and life sciences.

It's a post-trained version of Z.ai's GLM-5.3, roughly 753 billion parameters in a mixture-of-experts design. The company says it beats OpenAI's GPT-6-Astra on the FrontierScience-Research benchmark, 38.3% to 33.3%, and lands near Kimi-K3 on GPQA-Diamond. Those are the company's own numbers.

What stands out is the training data: electronic lab notebook records and discovery trajectories that include failed attempts, not just published papers. That's the kind of data that could make a model more useful at the bench.

We like this release. Science benefits from tools that others can inspect, rerun and argue with, and an open, permissively licensed model aimed at real lab work moves in that direction. Training on failed experiments is especially smart, because published literature mostly records what worked.

Now the caveats, and they matter. The benchmark wins are self-reported, and the internal evaluations are the company's own design. It also trails GPT-6-Astra on NMR structure elucidation. On the experiment-planning test, GPT-6-Astra and Opus-5.5 were left out because they refused more than half the requests. That tells us something about how closed models handle lab-planning questions, but it also means that comparison isn't head-to-head.

The refusal detail deserves a second look. If a model answers lab-planning questions that frontier labs decline, some of that may be overcaution on their side. Some of it may be caution that exists for a reason. Chemistry and life-science planning is dual-use territory, and an open model with no gatekeeper can't be recalled. We haven't seen the safety evaluation for Hydrogen-1.0, and we'd want to.

The word reproducible also needs a footnote. The weights are open, but the full datasets and training recipes aren't. Researchers can inspect the outputs and fine-tune, but they can't fully rebuild what was done.

Our read: a genuinely useful open tool for computational chemistry and materials work, with numbers to verify and a dual-use question the release doesn't answer yet.

The technical post shows the model writing and debugging a polymer simulation in GROMACS, mapping a reaction network, ranking candidate products from an NMR trace, and drafting a protocol for a 12-vial coupling study. Weights are on Hugging Face and ModelScope, tested on 8 GPUs with SGLang, with lower-precision builds for NVIDIA B200/B300 and AMD MI350X/MI355X.