Deep Principle Releases Hydrogen-1.0, an Open-Weight Scientific Model Derived from GLM-5.3
Startup Deep Principle has released Hydrogen-1.0, a model for chemistry, materials science, and life sciences built by post-training Z.ai’s GLM-5.3 on its own lab notebooks. The weights are under an MIT license; the benchmark scores, including an edge over GPT-6 Astra, come directly from the developer.
Startup Deep Principle released Hydrogen-1.0 on November 2, a large language model specialized in chemistry, materials science, and the life sciences. It is not a model trained from scratch: the team took GLM-5.3, Z.ai’s open model, and post-trained it on its proprietary laboratory data. The weights are published under an MIT license on Hugging Face and ModelScope. According to the developer, Hydrogen-1.0 outperforms OpenAI’s GPT-6 Astra on a scientific research benchmark—a claim that has not yet been independently verified.
Lab Notebooks as Training Data
The claimed originality lies in the data. In addition to knowledge in chemistry, materials, and biology, Deep Principle used its teams’ electronic laboratory notebooks: what the researcher sought to learn, the protocol followed, the observations, and the outcomes, including failures. To this, it added “discovery trajectories”—sequences of choices and outcomes that show how a research campaign advances. The training combines supervised learning and reinforcement learning guided by research results.
The blog also outlines a training infrastructure that runs across both Nvidia (CUDA) and AMD (ROCm) GPUs, built on Megatron and SGLang. According to the Hugging Face model card, the model supports a 1-million-token context window and up to 128,000 output tokens. The repository indicates a model of approximately 753 billion parameters, tested on 8 GPUs: this is not a model designed to run on a workstation. Four-bit quantized versions are provided for Nvidia Blackwell (NVFP4) and AMD Instinct MI350X and MI355X (MXFP4) chips. The full datasets and training recipes have not been published.
What the Developer’s Scores Show
According to Chenru Duan of Deep Principle, Hydrogen-1.0 achieves 38.3% on FrontierScience-Research, compared to 33.3% for GPT-6 Astra, and 93.0% on GPQA Diamond. The blog notes that this latter score is roughly on par with Kimi K3, Moonshot’s open model. Across three internal scenarios, according to him, the model outperforms GPT-6 Astra at maximum reasoning effort on scientific computing (76.9% vs. 71.8%) and reaction prediction (90.5% vs. 85.7%), but falls short on NMR structure determination (26.7% vs. 30.0%).
The developer also highlights a materials discovery benchmark: over five turns, the model proposes transition metal complexes and receives feedback on the outcomes of its own choices. By the fifth turn, the average best HOMO-LUMO gap reaches 4.318 eV for Hydrogen-1.0, compared to 3.689 eV for GPT-6 Astra and 3.353 eV for Kimi K3. These evaluations are internal, and their test sets are not public.
Where to Use It
The model is available through four channels: weights downloadable on Hugging Face (after agreeing to share account information) and ModelScope, an API via training partner NeuroSpark, and within Mira, Deep Principle’s “AI Scientist.” The model card cautions that its outputs may contain errors in facts, calculations, or citations, and should be verified before any decision is made.
Our take
Hydrogen-1.0 highlights a clear trend: open Chinese foundation models like GLM-5.3 are now serving as starting points for domain specialists layering their proprietary data on top. Deep Principle’s argument is sound in principle: lab notebooks, complete with their failures, are rare data that broad-scope frontier labs lack. Yet the evidence remains self-reported. The win over GPT-6 Astra hinges on a single public benchmark, the margin on GPQA Diamond is narrow, and the strongest results stem from internal evaluations. The note regarding experiment planning also warrants scrutiny: GPT-6 Astra and Claude Opus 5.5 were excluded from that comparison because they refused the majority of prompts, reflecting the guardrails of closed models as much as Hydrogen-1.0’s strengths. Finally, the MIT license is a major asset for research labs, but the model’s sheer footprint means local deployment is reserved for teams with multi-GPU clusters.
Sources
- Deep Principle — Introducing Hydrogen-1.0 (tech blog) Primary source
- Hugging Face — deep-principle/Hydrogen-1.0 model card Primary source
- ModelScope — Hydrogen-1.0 Primary source
- Deep Principle on X — Hydrogen-1.0 announcement Primary source
- Chenru Duan (Deep Principle) on X — reported scores Primary source
- Chenru Duan (Deep Principle) on X — internal scenario scores Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
