Hydrogen-1.0: Deep Principle releases open-weight 753B science model

Hydrogen-1.0 is an open-weight, 753B-parameter mixture-of-experts scientific model released by Deep Principle on Nov. 2, 2026, and post-trained from GLM-5.3 on chemistry, materials science and life sciences data.

Deep Principle introduced Hydrogen-1.0 on Nov. 2, 2026, as an open-weight model designed for scientific research. The company said it built the model by post-training GLM-5.3 on domain data from chemistry, materials science and life sciences, along with records and agent trajectories from Deep Principle's own laboratories. The weights are distributed under an MIT license through a gated Hugging Face repository, with additional routes through ModelScope, the NeuroSpark API and the Mira AI Scientist application.

The release pairs a large-scale architecture with a benchmark claim. Deep Principle reported that Hydrogen-1.0 scored 38.3% on FrontierScience-Research, compared with 33.3% for GPT-6-Astra. Chenru Duan, founder and chief technology officer of Deep Principle, posted the comparison on X as part of the release announcement.

Why is an open-weight scientific model significant now?

Hydrogen-1.0 is aimed at scientific domains rather than general-purpose chat or coding tasks. The model's stated purpose is to support chemistry, materials science and life sciences research, where models are often evaluated on benchmarks such as FrontierScience-Research and GPQA-Diamond. Deep Principle's decision to release weights rather than a closed API is central to its stated mission of democratized science.

The release also reflects a growing division of labor in model development. Deep Principle built Hydrogen-1.0 by adapting Z.ai's GLM-5.3, a base model it credited in the announcement. That arrangement lets a specialist organization contribute domain expertise while relying on another lab's foundation model.

What is Hydrogen-1.0 and how is it built?

Hydrogen-1.0 is a mixture-of-experts model with roughly 753 billion total parameters and about 40 billion active parameters, according to the Hugging Face repository. The repository lists 753,329,940,480 parameters in safetensors format and identifies GLM-5.3 as the base model. Deep Principle's launch post says the post-training data spans chemistry, materials science and life sciences and includes internal lab records and discovery trajectories.

The model supports a 1 million-token context window and a maximum output of 128,000 tokens. Deep Principle said the model can be served on eight GPUs with SGLang and supports NVIDIA CUDA and AMD ROCm. Those specifications are designed to make a model of this size practical for research groups with dedicated GPU infrastructure.

Specification	Detail
Total parameters	753,329,940,480 in safetensors (~753B)
Active parameters	~40B
Architecture	Mixture-of-experts
Base model	GLM-5.3 (Z.ai)
Context window	1 million tokens
Maximum output	128,000 tokens
License	MIT (gated repository)
Quantized variants	NVFP4 and MXFP4 on ModelScope
Serving	Eight GPUs via SGLang; NVIDIA CUDA and AMD ROCm

What do the benchmark results show?

Deep Principle's headline result is 38.3% on FrontierScience-Research, which the company used as its primary science benchmark. The same announcement compared that score with 33.3% for GPT-6-Astra. Both figures are self-reported by Deep Principle and appeared in the X post announcing the release.

Duan also reported a 93.0% score on GPQA-Diamond, a graduate-level science reasoning benchmark. The GPQA-Diamond result was included in the same X post as the FrontierScience-Research comparison. On hard research benchmarks, absolute scores can be more informative when paired with a reference model.

How was the model trained on laboratory data?

Deep Principle said training combined supervised fine-tuning and reinforcement learning on research outcomes. The company named Megatron and SGLang as the frameworks used in the process. The training data included discovery trajectories from its laboratories, according to the launch post.

Supervised fine-tuning aligns a model with labeled examples, while reinforcement learning optimizes behavior against a reward signal. The combination is meant to improve performance on tasks where the correct answer is visible only after experimental execution. Deep Principle's description points to a model that learned from operational research processes, not only from published results.

How does post-training from GLM-5.3 differ from building a model from scratch?

Post-training starts with an existing foundation model and adapts it to new domains, rather than training a new model on trillions of tokens of raw text. Deep Principle's approach uses GLM-5.3 as the base and adds domain-specific data through supervised fine-tuning and reinforcement learning. That reduces the data and compute needed to create a specialized model.

The trade-off is that the final model inherits the strengths and limits of the base model. Hydrogen-1.0's behavior will reflect GLM-5.3's underlying capabilities as well as Deep Principle's scientific post-training.

Who can use Hydrogen-1.0 and through which channels?

Hydrogen-1.0 is available as open weights under an MIT license, which permits broad commercial and academic use. The Hugging Face repository is gated, so users must request access before downloading. Deep Principle also released NVFP4 and MXFP4 quantized variants on ModelScope, which reduce the memory footprint for deployment.

1. Download the gated weights from the Hugging Face repository deep-principle/Hydrogen-1.0.
2. Access NVFP4 or MXFP4 quantized variants on ModelScope.
3. Call the model through the NeuroSpark API.
4. Use the Mira AI Scientist application for agent-based scientific workflows.

The serving requirement is eight GPUs with SGLang, with support for NVIDIA CUDA and AMD ROCm. That makes the model accessible to research organizations that own or rent multi-GPU nodes. The 1 million-token context window supports long scientific documents, experimental protocols and multi-step analyses.

Access through the NeuroSpark API and the Mira AI Scientist application extends the model beyond self-hosted deployment. Those routes are aimed at teams that want the model's capabilities without operating the infrastructure.

Why does a 1 million-token context matter for science?

Deep Principle lists a 1 million-token context window for Hydrogen-1.0, which lets the model process entire experimental protocols, long literature corpora or multi-step agent logs in a single pass. For scientific work, long-context handling can reduce the need to chunk documents and can preserve relationships across distant sections of a paper.

The 128,000-token maximum output supports generated artifacts such as synthesis procedures, crystal-structure descriptions or detailed experimental plans. Those capabilities are directly relevant to the chemistry, materials science and life sciences domains named in the launch post.

Why is Deep Principle releasing weights under an MIT license?

The license choice reflects Deep Principle's stated position on how scientific AI should be distributed. Duan said in the release that science should be open, collaborative and democratized rather than proprietary. He argued that only open weights let researchers inspect, adapt and build on AI advances. The MIT license is permissive, but the gated repository gives Deep Principle a control point for distribution.

We’re introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3 (@Zai_org) on our internal scientific data, lab records, and discovery trajectories. It scores 38.3% on FrontierScience-Research vs GPT-6-Astra’s 33.3% and scores 93.0% on GPQA-Diamonds. We believe science should be open, collaborative, and democratized, rather than proprietary. Only open weights let researchers inspect, adapt and build on the unprecedented speed of AI growth.Chenru Duan, Founder and CTO, Deep Principle

Deep Principle's framing positions Hydrogen-1.0 against closed systems such as GPT-6-Astra, which the company cited as a benchmark reference. The company's argument is that scientific progress depends on access to model internals, not only on API outputs.

An MIT license also lowers legal barriers for downstream use. Research groups can integrate the model into proprietary pipelines, and commercial labs can build products on top of it without negotiating separate agreements.

What does Hydrogen-1.0 mean for the broader frontier-model market?

The release is notable for its scale. Few open-weight releases reach roughly 753 billion parameters, and fewer are specialized for science. Hydrogen-1.0 is built on GLM-5.3, which Deep Principle credited to Z.ai in the announcement. That arrangement separates base-model development from domain adaptation.

For chemistry, materials science and life sciences teams, Hydrogen-1.0 offers an alternative to closed scientific AI services, assuming they can meet the hardware requirements. The model's context window and quantized variants lower some deployment barriers, though the eight-GPU requirement remains significant.

The benchmark comparison with GPT-6-Astra places Deep Principle in direct competition with a closed frontier system. That comparison invites independent evaluation, which is how open-weight claims are typically tested.

What are the limits of open-weight scientific models?

Open weight does not mean fully open. The Hugging Face repository is gated, and the post-training data includes Deep Principle's internal lab records, which are not part of the weight release. Researchers can inspect the model's parameters, but they cannot inspect the proprietary data used to create it.

The 38.3% and 93.0% figures come from Deep Principle's own announcement. Quantized variants can behave differently from the full-precision model on precise scientific tasks, so the NVFP4 and MXFP4 releases may need separate evaluation.

The MIT license applies to the weights, not to the base model's original training data or to the lab records Deep Principle used in post-training. Users should verify what the repository actually distributes before assuming full reproducibility.

What should research teams evaluate before adoption?

Independent evaluation of Hydrogen-1.0 should begin with the benchmark claims. Deep Principle's published scores are self-reported, so teams should rerun FrontierScience-Research and GPQA-Diamond tasks under their own prompts and settings before relying on them. The model's behavior on proprietary in-house datasets will matter more than its performance on public benchmarks.

Deployment decisions should also account for quantization. NVFP4 and MXFP4 variants reduce hardware requirements, but compressed weights can lose precision on tasks involving chemical structures or numerical reasoning. Teams should compare quantized outputs against the full model before adopting a lower-precision version in production.

What comes next for Deep Principle and Hydrogen-1.0?

Deep Principle described Hydrogen-1.0 as its first open-weight model for scientific research. The company's access channels include both a developer API and an application aimed at agent-based work, which indicates continued investment in scientific agents. The launch materials do not include a detailed roadmap.

The next test is external. Research groups can now request access, run the model on their own scientific tasks and compare its behavior with closed alternatives. Those results, rather than the announcement, will determine whether Hydrogen-1.0 changes how scientific AI is built.