Deep Principle
Deep Principle releases Hydrogen-1.0, an open model for scientific research
Post-trained from GLM-5.3 on laboratory records and discovery work, it writes and runs simulations, and its maker reports wins over GPT-6 Astra on several science tasks.
Deep Principle introduced Hydrogen-1.0 on 2 November 2026, an open-weights model for scientific problem-solving in chemistry, materials science and the life sciences. It is a post-trained version of GLM-5.3, the open model from Z.AI, with about 753 billion parameters, a one-million-token context window and up to 128,000 tokens of output.
Deep Principle trained it with supervised fine-tuning followed by reinforcement learning, using domain data, feedback from scientists, laboratory records and the trajectories of earlier discovery work. The weights are under the MIT licence; the training data and full recipe are not released.
What Deep Principle reports
- It outperforms GPT-6 Astra on FrontierScience-Research, a benchmark of open-ended research problems, and improves on GLM-5.3 on GPQA-Diamond.
- On internal tests it leads GPT-6 Astra at maximum thinking by 5.13 points on scientific computing and 4.76 points on one-step reaction prediction. GPT-6 Astra leads by 3.33 points on NMR structure elucidation.
- In a five-round design task for transition metal complexes, its best candidates reached a mean HOMO–LUMO gap of 4.318 eV, against 3.689 eV for GPT-6 Astra.
- For experiment design, GPT-6 Astra and Claude Opus 5.5 were left out of the comparison because they rejected more than half of the requests.
The demonstrations are practical. In one, the model writes Python to build a polyethylene melt, runs it in the GROMACS simulation package, hits a topology error, fixes it, restarts the run and analyses a 40-nanosecond trajectory for density, packing, diffusion and chain dimensions. In another, it writes a 12-vial protocol for a Suzuki coupling reaction, catches an aspiration error during a virtual rehearsal and recompiles the protocol.
Model outputs are supporting material for users to review.
Hydrogen-1.0 is available as open weights on Hugging Face (after an access request) and ModelScope, with quantised builds for NVIDIA Blackwell and AMD Instinct MI350-series GPUs. It can also be used through an API from NeuroSpark and inside Deep Principle’s Mira platform.
What it means for your business
Most businesses aren’t running molecular simulations, but the pattern applies widely: take a capable open model and post-train it on your own domain records until it beats a larger general model on the work you actually do. Deep Principle’s results suggest the specialised route can outperform a frontier model on narrow, well-defined tasks.
Its model card also sets the right expectation for any technical work: the output is material for a qualified person to check. An agent can run the simulation, read the spectrum and draft the protocol; the scientist or engineer decides what goes into the report. That division of labour lets experts spend their hours on judgement instead of setup.
Sources
- Introducing Hydrogen-1.0 (opens in a new tab) · Deep Principle
- Hydrogen-1.0 model card (opens in a new tab) · Deep Principle on Hugging Face
- Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery (opens in a new tab) · MarkTechPost
General information only, not legal, financial or professional advice. Check the sources for your circumstances.