Hydrogen-1.0
Deep Principle's 753B science model, post-trained from GLM-5.3: MIT license, 1M context, 4-bit builds for Blackwell and AMD.
In short. Hydrogen-1.0 is a very large AI model tuned for science. Deep Principle took Z.ai's GLM-5.3 and trained it further on chemistry, materials and life-science data and on records from its own laboratories, so that it can explain derivations, plan experiments and help with open research questions. The weights are free to use under the MIT license, but you need a server with several high-end GPUs, and the company's results come from its own tests.
Summary
Deep Principle, a Chinese AI-for-materials company, published Hydrogen-1.0 in a blog post dated 2 November 2026, and its Hugging Face repositories were created on 31 October. It is a text-only model built by post-training GLM-5.3, the 753B-parameter mixture of experts from Z.ai, on domain data from chemistry, materials science and life sciences plus records and agent trajectories from Deep Principle's own laboratories. Training combined supervised fine-tuning and reinforcement learning, built on Megatron and SGLang, with UniPat, Mind Lab and NeuroSpark named as partners. The total parameter count matches GLM-5.3 exactly, and the card lists a 1M-token context and a 128K-token maximum output. Deep Principle does not state the active parameters, but the architecture is GLM-5.3's, which activates 40B per token. The chat template defaults to the maximum reasoning effort, and low and high are also available.
The results are Deep Principle's own and mostly appear 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 scores higher than GPT-6 Astra on scientific computing and one-step reaction prediction at each of three thinking settings, by 5.13 and 4.76 percentage points against Astra at maximum thinking, while Astra stays 3.33 points ahead on NMR structure elucidation. In a test that designs transition-metal complexes over five rounds, the mean best HOMO-LUMO gap was 4.318 eV for Hydrogen-1.0, 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi K3. The company says absolute scores across tasks should not be compared, and the model card itself gives no scores.
There are three repositories. The BF16 weights hold 753B parameters in about 1,507 GB and sit behind Hugging Face's automatic access gate, while the NVFP4 build (about 464 GB, for NVIDIA Blackwell, tested on 8 B300 GPUs with SGLang and a 262,144-token context) and the MXFP4 build (about 423 GB, for AMD Instinct, tested on MI350X with ATOM at tensor parallel 4) are open. The BF16 example serves the model with SGLang on 8 GPUs and the GLM reasoning and tool-call parsers, and says to keep generation_config.json, because generation does not stop at the end of an answer without it. In the MXFP4 test, 128 concurrent long reasoning requests filled the KV cache and answer quality dropped sharply, while 48 stayed normal. The weights are MIT licensed, Deep Principle's own code and documentation use Apache 2.0, and third-party files keep their licenses. The base model GLM-5.3 has Z.ai's own license, so read the LICENSING.md and NOTICE files in the gated repository before commercial use. On 11 November 2026 Hugging Face showed about 40,500 downloads in 30 days and 289 likes.
Good for
- Scientific questions, derivations and open-ended research problems in chemistry, materials science and life sciences, with expert review of the answers
- Research agents that call tools through the GLM parsers, with a 1M-token context
- Labs and companies with a multi-GPU node that want to host an MIT-licensed 753B model, on 8 Blackwell GPUs with NVFP4 or on AMD Instinct with MXFP4
- Studying how supervised and reinforcement-learning post-training specialize an open foundation model
Watch out for
- All benchmark claims come from Deep Principle and mostly appear as charts, the card gives no scores, and the datasets and training recipes are not released
- It is text-only, and its outputs are not verified, so the card says answers may contain wrong facts, calculations or references and must not drive consequential decisions
- The BF16 weights are about 1,507 GB and gated, and even the 4-bit builds are about 423 to 464 GB
- The 4-bit builds can answer differently from BF16, and the NVFP4 example caps the context at 262,144 tokens instead of the stated 1M
- The MXFP4 test saw quality collapse once 128 long requests filled the KV cache, so size the cache for your concurrency
- The base model's license differs from the MIT label, so check the LICENSING.md and NOTICE files before you build a product
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Hydrogen-1.0-MXFP4 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Hydrogen-1.0 (BF16) BF16 | 753B 40B active | 1.73 TB | 866 GB | 520 GB | GPU server |
| Hydrogen-1.0-NVFP4 NVFP4 | 753B 40B active | 1.73 TB | 866 GB | 510 GB | Big-memory desktop |
| Hydrogen-1.0-MXFP4 MXFP4 | 753B 40B active | 1.73 TB | 866 GB | 465 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is Hydrogen-1.0?
- Hydrogen-1.0 is a very large AI model tuned for science. Deep Principle took Z.ai's GLM-5.3 and trained it further on chemistry, materials and life-science data and on records from its own laboratories, so that it can explain derivations, plan experiments and help with open research questions. The weights are free to use under the MIT license, but you need a server with several high-end GPUs, and the company's results come from its own tests.
- Can I use Hydrogen-1.0 commercially?
- Yes. Hydrogen-1.0 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Hydrogen-1.0 open source?
- Hydrogen-1.0 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Hydrogen-1.0 on my own computer?
- Only on a machine with a lot of memory. The smallest variant, Hydrogen-1.0-MXFP4, needs about 465 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Hydrogen-1.0 released?
- Hydrogen-1.0 was released by Deep Principle on November 2, 2026.
- How long a context window does Hydrogen-1.0 have?
- Hydrogen-1.0 supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Hydrogen-1.0 model card (Hugging Face) huggingface.co
- Introducing Hydrogen-1.0 (Deep Principle blog, 2 November 2026) blog.deepprinciple.com
- Hydrogen-1.0-NVFP4 model card (Hugging Face) huggingface.co
- Hydrogen-1.0-MXFP4 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| 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 | |
| AREX-2 BAAI | 27.4B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU |