Back home

Hydrogen-1.0

x 0.90
Visit source

Intro

Hydrogen-1.0 is the first open-weight scientific research model from Deep Principle, a Hangzhou-based AI-for-science startup, released on November 2, 2026. It is built by post-training Z.ai's GLM-5.3 on the company's internal scientific data, full-cycle lab records (experiment design, procedure, observations and outcomes) and materials-discovery trajectories, targeting chemistry, materials science and life sciences.

The weights are published under the MIT license on Hugging Face (deep-principle/Hydrogen-1.0, plus MXFP4 / NVFP4 quantized variants) with a 1M-token context window, and the model is also available inside Deep Principle's research platform MIRA (with usage rates as low as 0.1x from Nov 2 to Nov 30).

Explain

Key facts

Item Details
Publisher Deep Principle (announced by co-founder & CTO Chenru Duan)
Base model Z.ai GLM-5.3 (MoE, ~753B-parameter class)
Training data Internal scientific data, full-cycle lab records, discovery trajectories
Context 1M tokens
License MIT, open weights (BF16 / MXFP4 / NVFP4)
Access Hugging Face download, MIRA platform, API (priced near GLM-5.3)

Reported results (self-reported)

  • FrontierScience-Research: 38.3% vs GPT-6 Astra's 33.3% — claimed SOTA.
  • GPQA-Diamond: 93.0%, on par with the strongest models.
  • Transition-metal complex (TMC) discovery: mean best HOMO–LUMO gap of 4.318 eV in round 5 vs 3.689 eV for GPT-6 Astra.
  • On internal research scenarios, it scores higher than or on par with GPT flagships.

What makes it different

Most science models learn from papers and textbooks. Hydrogen-1.0 instead uses how scientists actually design, run, observe and conclude experiments as training signal, so it emphasizes experiment planning, simulation and materials screening decisions rather than just answer accuracy. Because the weights are open, outside researchers can rerun and check the vendor's claims.

Usage

How to use it

  1. Try it online — sign in to Deep Principle's research platform MIRA and pick Hydrogen-1.0; usage is discounted to as low as 0.1x until Nov 30.
  2. Download the weights — find deep-principle/Hydrogen-1.0 on Hugging Face (accept the access terms); pick the MXFP4 / NVFP4 variants if memory is tight, and serve with vLLM / SGLang builds that support the GLM-5.3 architecture.
  3. Typical use cases:
    • Chemistry & materials: screening candidate molecules/complexes, property reasoning, simulation planning;
    • Experiment planning: proposing the next round of experiments from past lab records;
    • Life sciences: joint reasoning over literature and data, using the long context for whole lab notebooks.
  4. Read the details — the official blog (blog.deepprinciple.com) and Chenru Duan's 13-post X thread cover training and evaluation.

Note: all scores are self-reported; evaluate on your own tasks before relying on it for research decisions.