Skip to content
Neural Daily

ModelsNews

Hydrogen-1.0: a 753B open-weight model for chemistry

🇭🇺 Olvasd el magyarul →

China's Deep Principle released Hydrogen-1.0, an MIT-licensed model tuned for chemistry, materials science and life sciences. Its own tests show it beating GPT-6-Astra on a research benchmark.

By: Mikola Máté Patrik · · 2 min read

Robotic arm pipetting into lab tubes on the Hydrogen-1.0 announcement image
Image: Deep Principle · Press kit · Official poster image from the Deep Principle blog post · Source

What happened?

Hangzhou-based Deep Principle released Hydrogen-1.0, an open-weight model for scientific research (Deep Principle). It is built by post-training Z.ai's GLM-5.3 on chemistry, materials science and life sciences data. The goal is an open model that helps scientists frame questions, plan experiments and interpret results.

How it was built

Training combines SFT and RL on four kinds of data: domain knowledge, task feedback from the company's scientists, full-cycle lab records captured through its Mira electronic lab notebook, and agent discovery trajectories, including failed attempts. The full datasets and recipes are not released (MarkTechPost).

Specs and serving

  • 753 billion parameter MoE, with about 40 billion active per the GLM-5.3 base.
  • 1 million token context, up to 128K tokens of output.
  • MIT license. On Hugging Face, downloading requires an access request, which is approved automatically.
  • Tested in BF16 with SGLang on 8 GPUs. NVFP4 builds target NVIDIA B200 and B300, MXFP4 builds target AMD MI350X and MI355X.
  • The server exposes an OpenAI-compatible API with a reasoning_effort setting.

Benchmarks

All numbers are self-reported. On FrontierScience-Research it scores 38.3% versus 33.3% for GPT-6-Astra. On GPQA-Diamond, one of the founders reports 93.0%, roughly matching Kimi-K3. On internal tasks it leads GPT-6-Astra at maximum thinking by 5.13 points in scientific computing and 4.76 points in forward reaction prediction, but trails by 3.33 points in NMR structure elucidation.

On 117 experiment-planning questions it scored above Kimi-K3. GPT-6-Astra and Opus 5.5 were excluded from that test after refusing more than half of the requests. The demos show the model writing and fixing a GROMACS simulation, growing a reaction network, and drafting a JSON protocol for a 12-vial Suzuki coupling study.

Who should care

For teams building chemistry or materials agents, this is one of the first large scientific models you can run yourself. The 8-GPU requirement puts it in research-institute or company infrastructure territory, though. Since the numbers are self-reported, independent tests are worth waiting for.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles