Open Science Model Hydrogen-1.0: First, Learn Not to Grab the Wrong Pipette Tip in a Virtual Lab

Deep Principle released Hydrogen-1.0, an open-weight model for scientific research. It is post-trained on GLM-5.3, with weights on Hugging Face under the MIT license. The official headline results are straightforward: FrontierScience-Research pass rate 38.3%, GPT-6-Astra 33.3%; 93.0% on GPQA-Diamonds. API cost is close to GLM-5.3.

The model naming starts with hydrogen. Some users on X joked that this is the first naming scheme that treats the periodic table as a roadmap, with 117 names still to come. Chenru Duan replied, "exactly." Jokes aside, Hydrogen-1.0 is positioned for scientific problem-solving and research assistance. Training data includes domain data, scientists' task feedback, lab records, and complete discovery trajectories. In the author's reply, complete discovery trajectories are explained as agent trajectories from projects.

A user asked how much of the 38.3% comes from post-training. Chenru Duan answered that GLM-5.3 itself scores 24.4% on FrontierScience-Research, so the starting point was already strong. This clarifies the value of the open-weight base model and shows that Hydrogen-1.0's gains indeed come from post-training and scientific data.

Comparisons on internal research scenarios are more granular. Scientific computing: 76.9% vs. GPT-6-Astra's 71.8%; forward reaction: 90.5% vs. 85.7%; NMR structure elucidation: 26.7% vs. 30.0%. Hydrogen-1.0 wins the first two and loses the NMR one. The official statement says these scenarios were tested in maximum thinking mode.

The computational chemistry example is very concrete. To study what a plastic bag looks like at the molecular scale, Hydrogen-1.0 builds a polyethylene melt, runs GROMACS, diagnoses and fixes topology errors, then analyzes the newly generated 40 ns trajectory and reports density, packing, diffusion, and chain dimensions.

The lab scenario better reveals where the training data comes from. Hydrogen-1.0's training covers full-cycle lab records: how scientists initially design an experiment, the operating steps, observed results, and final results. This allows it to do experiment planning even when given only a brief intent description. The authors mention that AI taking over a lab carries many risks, so Hydrogen-1.0 does virtual rehearsals. In the demo, it catches an attempt to aspirate liquid without changing to a new pipette tip, fixes the action order, checks again, and lets the simulation continue.

Failed experiment data is another detail. A user asked whether lab records include failed experiments, which are data rarely published in papers. Chenru Duan said yes, but frankly, they have not yet found a way for the model to learn much from this data, and benchmarking is also hard. This answer is more informative than the demo. It is not surprising for a scientific model to memorize successful paths; the hard part is learning from failure.

The goal in materials discovery is to be faster. The official claim is that in transition metal complex searches, Hydrogen-1.0 finds better candidates with fewer attempts, thereby accelerating the discovery process.

The model specs are also right here: 753B parameters, 1M token context, 128K token maximum output, BF16 unquantized weights. There are also an NVFP4 version for NVIDIA Blackwell B200 and B300, and an MXFP4 version for AMD Instinct MI350X and MI355X. Deployment was tested with SGLang on 8 GPUs. MIT license, with clear fine-tuning rights.

Limitations should also be stated clearly. The Hugging Face model card states that the full training data and recipe were not released with the weights; outputs require scientific review and cannot be treated as verified experimental conclusions. A user asked about the Terminal-Bench Science score. Chenru Duan said this benchmark is led by Steven Dillmann, that he was involved early on, that competing would be unfair, and that a future version will be released.

How to access: weights are on Hugging Face, API is on NeuroSpark, the lab agent is called Mira, and the technical blog is here.

The periodic-table naming scheme is clever; after hydrogen there are plenty of names left. What really matters is whether the next model can learn from failed experiments, and what the open-weight community will build with it.