Image: Deep Principle
Deep Principle, a startup focused on AI for chemistry and materials, has released Hydrogen-1.0, its first open-weight model for scientific research. The company describes it as a model built by post-training Z.ai's GLM-5.3 on chemistry, materials science and life sciences data, along with laboratory records and agent trajectories from its own labs.
Chenru Duan, listed on the company's website as founder and CTO, announced the model on X on November 2, saying it scores 38.3% on FrontierScience-Research against 33.3% for GPT-6-Astra and 93.0% on GPQA-Diamond.
We’re introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3 (@Zai_org) on our internal scientific data, lab records, and discovery trajectories.
— Chenru Duan (@chenru_duan) November 2, 2026
The weights are available on Hugging Face and ModelScope under the MIT License, with separate NVFP4 and MXFP4 quantized versions for NVIDIA Blackwell and AMD Instinct hardware. The model card lists a context length of one million tokens and a maximum output of 128,000 tokens, and states that the full training datasets and recipes are not released with the weights. Hydrogen-1.0 can also be used through an API on NeuroSpark and inside Mira, the company's AI scientist product.
According to the official Deep Principle blog, the model outperforms GPT-6-Astra on FrontierScience-Research and improves on GLM-5.3 on GPQA-Diamond, roughly matching Kimi-K3. Deep Principle says it reaches these results at an API cost similar to GLM-5.3, since it uses that model's pricing. All of the figures come from the company's own evaluations, and no independent results had been published at the time of this writing.
Image: Deep Principle
Deep Principle also reports results from internal research scenarios. Against GPT-6-Astra at maximum thinking, Duan said Hydrogen-1.0 scored 76.9% versus 71.8% on scientific computing and 90.5% versus 85.7% on forward reaction prediction, while trailing on NMR structure elucidation with 26.7% against 30.0%. The blog notes that differences across these internal tasks reflect the difficulty of the cases and should not be read as direct comparisons of model capability.
The training pipeline combines supervised fine-tuning and reinforcement learning. Deep Principle says it draws on domain data, task feedback from its scientists, full cycle laboratory records covering intent, procedure, observations and outcomes, and complete discovery trajectories that include failed attempts. Duan said in a reply on X that the lab records come from Mira, which the company uses in place of a conventional electronic lab notebook.
The blog shows several worked examples. In one, the model writes Python to build a polyethylene melt, runs GROMACS, repairs a topology error and analyzes a 40 ns trajectory. In another, it writes a protocol for a 12 vial Suzuki coupling study, and a virtual rehearsal catches an attempt to aspirate without a fresh pipette tip before the corrected protocol is compiled.
In a materials discovery test, models recommended transition metal complexes over five rounds, each seeing only the results of its own picks. Deep Principle reports a mean best so far HOMO LUMO gap of 4.318 eV for Hydrogen-1.0 after round five, compared with 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3.
Image: Deep Principle
The model card cautions that outputs may contain incorrect facts, calculations or references and should not be treated as verified experimental findings. Coverage from outlets such as TLDR has so far repeated the company's benchmark claims. Deep Principle frames the release as a push for open scientific AI, arguing that only open weights let researchers inspect, adapt and build on the models behind discovery.
Comments