Hydrogen-1.0: a 753B open science model built on GLM-5.3

Deep Principle released Hydrogen-1.0, a 753B science model post-trained from GLM-5.3 under an MIT label, with 1M context and 4-bit builds. All benchmark claims are its own.

  • New release
  • 2 min read
  • Published

Deep Principle, a Chinese AI-for-materials company, published Hydrogen-1.0 in a blog post dated 2 November 2026. It is a text-only model made by post-training GLM-5.3, the 753-billion-parameter model from Z.ai, on chemistry, materials-science and life-science data and on records from Deep Principle’s own laboratories. The weights carry an MIT label.

What is new

Training combined supervised fine-tuning with reinforcement learning, so the parameter count matches GLM-5.3 exactly: 753 billion, in a mixture-of-experts design. Deep Principle does not state the active parameters, but the base model activates 40 billion per token. The card lists a 1M-token context and a 128K-token maximum output, and the reasoning effort defaults to maximum, with low and high available.

The results are Deep Principle’s own and mostly come as charts. The blog says Hydrogen-1.0 beats GPT-6-Astra on FrontierScience-Research and improves on GLM-5.3 on GPQA Diamond, roughly matching Kimi K3. On internal tasks it leads Astra at maximum thinking by 5.13 points on scientific computing and 4.76 on one-step reaction prediction, while Astra stays 3.33 points ahead on NMR structure elucidation. The model card gives no scores.

Running it

There are three repositories. The BF16 weights are about 1,507 GB and sit behind Hugging Face’s automatic access gate. The open NVFP4 build is about 464 GB and was tested on 8 NVIDIA B300 GPUs with SGLang at a 262,144-token context. The open MXFP4 build is about 423 GB and was tested on AMD MI350X GPUs with ATOM. In that test, 128 concurrent long reasoning requests filled the KV cache and answer quality dropped sharply, while 48 stayed normal. For the BF16 weights, keep generation_config.json, because generation does not stop at the end of an answer without it.

License

The card labels the weights MIT, Deep Principle’s own code and documentation Apache 2.0, and third-party files keep their licenses. The base model uses the GLM-5.3 License, which is close to MIT but adds a security review for model-as-a-service operators above US$10 billion in revenue. The gated repository holds LICENSING.md and NOTICE files that we could not read, so check them before you build a product.

Why it matters

Hydrogen-1.0 is the only open model in our catalog tuned for science, and it comes from a company that runs its own laboratories. Like Ornith 1.5 for coding, it shows a company post-training a large open base model for its own field and releasing the result under a simple label.

The limits are just as clear. Even the 4-bit builds need 423 to 464 GB, the datasets and training recipes are not released, and the card warns that answers may contain wrong facts, calculations or references. Treat it as a model to test on your own chemistry tasks with expert review. The best open-source LLMs list places it among the other giant models.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Hydrogen-1.0 Deep Principle 753B 40B active 1M MIT (Permissive) 465 GB Big-memory desktop
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
Kimi K3 Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
Ornith 1.5 DeepReinforce 9.65B – 403B 17B active 256K MIT (Permissive) 6.66 GB Laptop or phone