Hydrogen-1.0: a scientific model that repairs code and writes laboratory protocols

In a Hydrogen-1.0 demonstration, a polyethylene calculation stops because of an error in the description of the molecular system. The model fixes the file and reruns the calculation, then analyses the resulting molecular trajectory. This is a concrete scientific-AI outcome: a calculation prepared and repaired by the model, then performed by external software. Several separate checks still stand between that result and an independent scientific discovery.

Deep Principle introduced Hydrogen-1.0 on November 2, 2026. The team’s technical account describes the scientific demonstrations, post-training data and internal comparisons. The model cards specify the available weight formats and deployment requirements.

Hydrogen-1.0 is a 753-billion-parameter language model created by Deep Principle through further training of Z.AI's GLM-5.3. Parameters are numerical values adjusted during training to improve prediction and response generation. The Hydrogen-1.0 model card describes text-based interaction: scientific questions, formula derivations and analysis of research tasks. The published weights allow the model to run in a user's own computing environment.

In the polyethylene example, the model writes Python code to prepare a melt and launches GROMACS, molecular-dynamics software that calculates atomic motion. The error concerns topology: the description of how molecules are structured in the simulated system. After fixing it, the model restarts the simulation and analyses a 40-nanosecond trajectory. Deep Principle lists the properties examined: density, molecular packing, diffusion and polymer-chain dimensions. These actions are described in the technical account of the demonstration.

The division of labour matters when assessing the result. The model writes code, selects actions and handles errors; GROMACS performs the physical calculation. A successful restart shows that the system continued after a failure in this example. Whether the simulation supports a scientific conclusion also depends on the problem formulation and computational assumptions. The published demonstration does not establish the physical validity of every simulation Hydrogen might prepare.

In another example, the model investigates a reaction network involving the Ireland–Claisen rearrangement. Deep Principle describes 27 molecular structures and 30 plausible reactions, including competing pathways and reversals. Such a result helps compare possible routes through which a substance might change. “Plausible” remains an important qualification: the proposed network requires checking, and calculating energy barriers alone does not establish that all the listed substances have been produced in a laboratory.

The next demonstration starts with seven possible products of a 2-bromonaphthalene reaction. They differ in the position at which a formyl group, CHO, is attached. The model compares the candidates with a supplied proton NMR spectrum—a record of hydrogen-nucleus signals that chemists use to infer the environment of atoms in a sample.

Hydrogen examines downfield signals, located at larger values on the spectrum's scale, and the crowded region associated with aromatic rings, then assesses the proposed structures. Many signals in the latter region lie close together, complicating the comparison. In the demonstration itself, the product assignments are labelled preliminary. Spectrum analysis and assessment of candidates therefore describe the result more accurately than a definitive determination of the sample's composition.

The model card separately warns that Hydrogen can make mistakes in facts, calculations, equations, references and scientific explanations. A detailed answer is not confirmation of accuracy. That qualification also applies to a convincing account of a spectrum: the model's response remains material for scientific verification.

The laboratory-protocol demonstration makes the boundary between model and equipment particularly clear. The research objective is to investigate Suzuki coupling while using less palladium catalyst. Hydrogen prepares a JSON file with instructions for a 12-vial study. The sequence covers solid dispensing, liquid transfer, adding magnetic stir bars, running the reaction, preparing samples and handing them over for analysis.

Before execution, the instructions are checked in a virtual laboratory environment. It accounts for the workstation's capabilities, dependencies between operations and the conditions for individual actions. The check catches an attempt to draw liquid without a new pipette tip. The model corrects the sequence, after which the revised protocol passes validation. Deep Principle calls this mechanism lab compilation: a plan becomes instructions that can be checked against the configuration of a particular laboratory.

The interactive demonstration is labelled a programmed virtual replay, with no connection to operating equipment. It supports an account of the protocol-checking and correction sequence. It does not establish that an autonomous chemical experiment took place. Virtual validation found a specific error in instrument handling; actual execution and the reaction's scientific outcome require their own evidence.

Deep Principle says it used scientific data from chemistry, materials science and the life sciences, researcher feedback, laboratory records and complete sequences of research actions. Such records retain the connection between an objective, the selected procedure, observations and the subsequent decision. Training also includes failed candidates and revised hypotheses.

The first stage trains the model on prepared examples of reasoning, planning, protocol writing and result analysis. Reinforcement learning follows: feedback from research actions is used to change the model's behaviour. The authors intend this to help the system choose its next experiment in light of the previous outcome. The public account outlines the approach; the full training datasets and a reproduction recipe have not been released alongside the weights.

A separate test concerns transition-metal complexes. Models must propose compounds with a larger difference between the energies of the highest occupied and lowest unoccupied molecular orbitals—the HOMO–LUMO gap. Orbitals describe electron states in a molecule; this comparison concerns the highest-energy occupied state and the nearest unoccupied one. The energy gap is the specified numerical search target, rather than a general measure of material quality. All models start from the same initial set of compounds and make proposals over five rounds. Thereafter, each receives results only for its own proposals.

According to Deep Principle's data, by the fifth round the mean best value reached was 4.318 electronvolts for Hydrogen-1.0, 3.689 for GPT-6-Astra and 3.353 for Kimi-K3. This experiment compared the models' computational search for compounds with a specified property.

Editorial graphic based on Deep Principle's data, November 2, 2026. It shows the mean best HOMO–LUMO energy gap found by the fifth search round. The comparison concerns computational search. This report supplies no data on synthesis of the proposed compounds or independent reproduction of the result.

The Deep Principle technical account reports a lead over GPT-6-Astra on FrontierScience-Research and advantages of 5.13 and 4.76 percentage points on internal scientific-computing and reaction-product-prediction tasks, compared with Astra at its maximum reasoning setting. For NMR structure analysis, Astra at that setting instead leads by 3.33 points. These are evaluations by Hydrogen's developer, rather than independent tests by this publication.

Experimental planning was compared with Kimi-K3 on a shared set of 117 questions. GPT-6-Astra and Opus-5.5 were excluded from this part because they declined more than half the requests. The reason matters: the planning chart cannot rank all four models by the quality of the plans they produced.

Cost comparisons have a further qualification. The authors use GLM-5.3 pricing as a proxy for Hydrogen. Some charts count only output-token costs, including reasoning, excluding input data and answer evaluation. These charts compare performance with a particular estimate of API expenditure. They do not show the full cost of operating a Hydrogen server or consuming laboratory resources.

Three variants have been published: original BF16 weights and compressed NVFP4 and MXFP4 versions. The names identify numerical storage formats: compressed versions reduce weight size but require compatible software and hardware. When checked on November 8, the files in the NVFP4 repository totalled approximately 464.19 GB in decimal units. That is a file size, rather than a ready-made VRAM requirement: execution also needs memory for intermediate data, including saved results of processing earlier tokens—pieces of text. Requirements vary with input length and concurrent requests.

In the NVFP4 model card, Deep Principle reports testing on eight NVIDIA B300 accelerators with SGLang. For MXFP4, it specifies AMD Instinct MI350X, the ATOM environment and distribution across four accelerators. These are server configurations reported by the team, rather than a promise of compatibility with any graphics card from the relevant manufacturer. The stated context reaches one million tokens, but the supplied NVFP4 launch example is limited to 262,144 tokens.

According to the licensing section of Hydrogen's model card, all three weight variants are labelled MIT-licensed; Deep Principle's own code and documentation use Apache 2.0, while third-party files retain their respective licences. Hugging Face requires sign-in and consent to share account information with the publisher before downloading. “Open weights” here therefore means the ability to obtain and use the trained model under those terms. Fully repeating its training would also require data and a detailed procedure, which are absent from the release.

Hydrogen-1.0 already offers something concrete for external examination: available weights and described tasks with observable intermediate actions. In the GROMACS and Suzuki-protocol examples, readers can trace how an error message changes the model's next action. Open weights let other researchers examine such behaviour on their own tasks, but the laboratory part of the published demonstration currently ends with virtual instruction validation. It supplies no evidence of this protocol running on a real installation.