Skip to content
Chenru Duan@chenru_duan · Nov 2, 2026

We’re introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3…

7 tweets1 min read♥ 3.7Koriginal

Summary

Hydrogen-1.0 is an open-weight AI model designed for scientific research that outperforms proprietary models like GPT-6-Astra on chemistry and physics benchmarks. The model excels at experimental planning, molecular simulation, and materials discovery while prioritizing safety through virtual rehearsals before lab execution.

Summarized by ThreadOut AI from the full thread. May miss nuance — read the thread below.

  1. #1

    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. It scores 38.3% on FrontierScience-Research vs GPT-6-Astra’s 33.3% and scores 93.0% on GPQA-Diamonds. We believe science should be open, collaborative, and democratized, rather than proprietary. Only open weights let researchers inspect, adapt and build on the unprecedented speed of AI growth. There are many recent models named by chemical elements, but we believe Hydrogen-1.0 is the first aiming towards and open scientific superintelligence for discovery. 🧵

  2. #2

    Hydrogen-1.0 gives SOTA performance on FrontierScience-Research and on par to SOTA on GPQA-Diamond at an API cost similar to GLM-5.3. On our internal research scenarios, Hydrogen-1.0 scores higher than or on par with GPT-6-Astra at maximum thinking on scientific computing (76.9% vs 71.8%), forward reaction (90.5% vs 85.7%), and NMR structure elucidation (26.7% vs 30.0%). read the full blog at blog.deepprinciple.com/introducing-hy…

  3. #3

    How can it help computational chemists and materials scientists? To study a plastic bag in a molecular scale, Hydrogen-1.0 builds a polyethylene melt, runs GROMACS, diagnoses and repairs a topology error, then analyzes a fresh 40 ns trajectory: density, packing, diffusion and chain dimensions.

  4. #4

    We train on full-cycle lab records: how a scientist design an experiment in the beginning, the procedure, the observations, and the outcome. This information helps Hydrogen-1.0 excel in experiment planning under a short description of intent. But AI taking over the lab has tons of risks. Hydrogen-1.0 makes a virtual rehearsal that catches an attempt to aspirate without a fresh pipette tip. The action order is repaired and rechecked before the simulated run proceeds.

  5. #5

    All of the above leads to a model that discover new materials quicker. In our case of searching for transition metal complexes, Hydrogen-1.0 finds better candidates with much fewer attempts, thus accelerating the discovery itself.

  6. #6

    Science is infinite. We want people to contribute on building models that learn from new experiments and discoveries that everyone can use. Explore Hydrogen-1.0 and build on our open weights. Weights: huggingface.co/deep-principle… API: neuro-spark.ai/dashboard/mode… Mira: mira.deepprinciple.com Blog: blog.deepprinciple.com/introducing-hy…

  7. #7

    @Zai_org full video for people who want more glasses: youtu.be/dD5TBiZQGAE