Deep Principle's Hydrogen-1.0 Beats GPT-6 on Science With 753B Open Parameters
Deep Principle released Hydrogen-1.0 , a 753B open-weight scientific model under MIT license. Post-trained from GLM-5.3 on chemistry, materials science, life-science data, and internal lab records. Scores 38.3% on FrontierScience-Research vs GPT-6-Astra's 33.3%, and 93.0% on GPQA-Diamond.
Takeaways
- Deep Principle released Hydrogen-1.0 , a 753B open-weight scientific model under MIT license.
- Post-trained from GLM-5.3 on chemistry, materials science, life-science data, and internal lab records.
- Scores 38.3% on FrontierScience-Research vs GPT-6-Astra's 33.3%, and 93.0% on GPQA-Diamond.
- Ships in BF16, NVFP4 (Blackwell), and MXFP4 (AMD MI350X) with 1M context, 128K output.
- Beats GPT-6-Astra on scientific computing and forward reaction prediction; loses on NMR structure elucidation.
- Deployable via SGLang on 8 GPUs; API priced at GLM-5.3 rates. Tech blog here .
Deep Principle, a Hangzhou startup, has released Hydrogen-1.0 , a 753-billion-parameter model post-trained for chemistry, materials science, and life sciences. The company publishes the weights on Hugging Face under the MIT license, giving research groups a path to inspect, adapt, and run the model on their own infrastructure.
Co-founder and CTO Chenru Duan describes Hydrogen-1.0 as the company’s first open-weight research model. It starts from Z.ai’s GLM-5.3 and adds training on internal scientific data, laboratory records, and discovery trajectories. Deep Principle reports a 38.3% pass rate on OpenAI’s FrontierScience Research evaluation, compared with 33.3% for GPT-6 Astra. Those results remain vendor-reported.
What 753 billion parameters require
Hydrogen-1.0 uses a mixture-of-experts architecture, which routes each token through a subset of specialized blocks instead of activating every parameter. Its glm_moe_dsa architecture tag indicates that it inherits GLM-5.3’s mixture-of-experts design and deep sparse attention. Deep Principle has not disclosed the number of active parameters per token, an important figure for estimating inference speed and compute requirements.
The advertised context window is one million tokens, with a maximum output of 128,000 tokens. Long contexts also expand the key-value cache used during inference, so practical memory requirements extend beyond the model weights.
| Package | Target hardware | Approximate raw weight size |
|---|---|---|
| BF16 | Hardware with sufficient 16-bit memory | About 1.5 TB |
| NVFP4 | NVIDIA Blackwell B200 and B300 GPUs | About 377 GB before metadata |
| MXFP4 | AMD Instinct MI350X and MI355X GPUs | About 377 GB before metadata |
The estimates cover parameter storage alone. Quantization scales, runtime buffers, the context cache, and framework overhead increase actual memory use. The Hugging Face repository is gated, so users must sign in and accept its terms before downloading the model shards.
Lab trajectories shape the training
Deep Principle trained Hydrogen-1.0 on complete experimental trajectories that include question framing, procedure design, observations, failed attempts, and subsequent revisions. That material can teach relationships between decisions and outcomes that finished papers often omit.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
