Hydrogen-1.0: a 753B open-weight model trained on lab records reports a science win over GPT-6-Astra

Deep Principle introduced Hydrogen-1.0 on November 2, 2026, a scientific model built by post-training GLM-5.3 on domain data from chemistry, materials science and life sciences, together with records and agent trajectories from its own laboratories. The Hangzhou company, which builds AI for materials research, describes the model as a research assistant for scientific problem-solving: framing questions, planning experiments, analysing results and pursuing open-ended discovery.

The pitch is reproducibility. Deep Principle notes that recent AI-assisted science results — it cites an OpenAI proof effort and an Anthropic enzyme discovery — came from proprietary models whose discovery capabilities are not available as open weights, and in some cases not through a public API at all. Hydrogen-1.0 is its attempt at an inspectable alternative: researchers can download the weights, test the methods and build on shared improvements, rather than taking a vendor's claims on trust.

Trained on the whole experiment, not just the papers

Training combines supervised fine-tuning with reinforcement learning, drawing on four sources: domain knowledge in chemistry, materials science and life sciences; task-specific feedback from Deep Principle's scientists; full-cycle laboratory records; and discovery trajectories — sequences of choices and feedback, including failed attempts. The laboratory records come from the company's electronic lab notebooks, documented through its Mira agentic system, preserving what a scientist wanted to learn, how the experiment was performed, what happened and what to investigate next.

That distinction is the bet. Papers record outcomes; lab records record the decisions that produced them, including the ones that failed. Deep Principle argues that learning from those trajectories teaches the model how research decisions lead to outcomes — which candidate to explore, which synthesis path to consider, which observation should change a hypothesis. The company has not released the full datasets or training recipes, MarkTechPost reports, so outside researchers cannot yet retrace exactly what the model was taught. Its model card also warns that outputs can contain incorrect facts, calculations, equations or references, and advises verifying scientific claims against primary literature, independent calculations or experiments.

What the numbers say — and what they do not prove

On public benchmarks, Deep Principle reports that Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research, scoring 38.3% against 33.3% for GPT-6-Astra, and improves over GLM-5.3 on GPQA-Diamond, where it reports 93.0% and roughly matches the open frontier model Kimi-K3. The GPQA-Diamond figure was reported from a thread by one of the founders, according to MarkTechPost. Those are the company's own evaluations; no independent benchmark result has been published.

Deep Principle's internal evaluations compare the model with GPT-6-Astra on three tasks its research team meets often: scientific computing, one-step forward reaction prediction and NMR structure elucidation. Against GPT-6-Astra at maximum thinking, Hydrogen-1.0 leads by 5.13 and 4.76 percentage points on the first two, while GPT-6-Astra leads by 3.33 points on NMR elucidation. On 117 experiment-planning questions, Hydrogen-1.0 scored above Kimi-K3; GPT-6-Astra and Opus-5.5 were omitted from that comparison after refusing more than half of the requests. The company itself cautions that differences in absolute scores across its internal tasks reflect how difficult those cases are, and should not be read as direct comparisons of overall model capability.

In one open-ended test, each model recommended new transition-metal complexes over five rounds, starting from the same initial population and learning only from its own recommendations. By round five, Hydrogen-1.0 reached a mean best-so-far HOMO–LUMO gap of 4.318 eV, against 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3. It is a narrow, company-run demonstration of discovery behaviour — promising for materials search, and still a long way from an independently validated result.

Four demos from the bench

The announcement's demonstrations show the model working through complete workflows rather than answering isolated questions. Hydrogen-1.0 writes Python to build a polyethylene melt, runs a GROMACS simulation, detects and repairs a topology error, and analyses the corrected 40-nanosecond trajectory. It grows an Ireland–Claisen rearrangement network of 27 structures through 30 plausible reactions, and ranks seven candidate monoformylation products of 2-bromonaphthalene from a supplied proton NMR trace. In a fourth demo, it writes a JSON protocol for a 12-vial Suzuki coupling study; a virtual lab-compilation step catches an attempt to aspirate without a fresh tip, and the model fixes the action order before the protocol compiles.

Those demos show what the intended workflow looks like: propose, simulate or rehearse, catch the error, revise. They do not by themselves establish that every scientific conclusion in them was correct — which is why the model card's warning matters. The value of an open model here is that other researchers can run the same workflows and check.

How to run it

Hydrogen-1.0 is a 753-billion-parameter mixture-of-experts model, with about 40 billion active parameters inherited from its GLM-5.3 base, a 1-million-token context window and a 128,000-token maximum output, according to MarkTechPost's summary of the release. The weights are available under the MIT licence on Hugging Face and ModelScope, with full-precision BF16 weights and quantised builds: NVFP4 for NVIDIA B200 and B300 GPUs, and MXFP4 for AMD Instinct MI350X and MI355X hardware. Deep Principle says it tested BF16 serving with SGLang across eight GPUs, and the server exposes an OpenAI-compatible API. The model is also available through an API from NeuroSpark and inside Mira, the company's AI Scientist product.

Whether Hydrogen-1.0 earns a place in real research workflows will be decided by independent testing: other groups rerunning its benchmark claims, probing its failure modes on their own problems, and checking whether lab-record training generalises beyond one company's notebooks. The weights being open makes that test possible. The scores being the company's own means it has not happened yet.