# Hydrogen-1.0: Science Benchmarks, Open Weights and Deployment Limits

Deep Principle's GLM-5.3-based science model has promising reported results, but its deployment details deserve as much attention as its scores.

By Paras Madan | 2026-11-04


Deep Principle introduced Hydrogen-1.0 on November 2, an open-weight scientific model built by post-training GLM-5.3 on scientific data, laboratory records and discovery trajectories. It targets research assistance in chemistry, materials science and life sciences, with reported gains that make it worth evaluating beyond general-purpose chat benchmarks.

The release also comes with a useful reminder: access to weights, a large context window and a benchmark lead describe different parts of a research system. None establishes that the system can independently produce reliable science.

## Hydrogen-1.0 specifications and access

The [Hugging Face model page](https://huggingface.co/deep-principle/Hydrogen-1.0) lists 753 billion parameters. Its BF16 model card specifies a one-million-token context length and a maximum output of 128,000 tokens. Actual budgets depend on the runtime, service settings and available resources.

The weights carry an MIT license. Deep Principle's independently authored code and documentation use Apache 2.0, while third-party material retains its respective licenses. The BF16 Hugging Face repository currently asks users to accept access conditions and share account information before accessing files.

Full training datasets and complete recipes are not included. Researchers can work with the resulting model, but the release does not provide everything needed to reproduce its training.

Deep Principle's technical blog lists four access routes: Hugging Face and ModelScope for weights, NeuroSpark for API access, and its Mira AI Scientist product. Hugging Face itself lists no deployed Inference Provider. That is a limitation of that hub's hosting options, not evidence that no hosted route exists elsewhere.

## What the science benchmarks show

In his [launch announcement](https://x.com/chenru_duan/status/2117249292469665838), founder and CTO Chenru Duan reported 38.3% on FrontierScience-Research, against 33.3% for GPT-6-Astra, and 93.0% on GPQA Diamond. These are Deep Principle's reported results, not an independent validation of a general scientific advantage.

The [technical report](https://blog.deepprinciple.com/introducing-hydrogen-1.0/) adds a more mixed picture. In internal comparisons against Astra at maximum thinking, Hydrogen leads by 5.13 percentage points on scientific computing and 4.76 points on forward reaction prediction. Astra leads by 3.33 points on NMR structure elucidation.

Those differences matter when choosing a model for a particular workflow. A lab working on structure analysis should not infer the best fit from an aggregate science headline.

There is also a pricing caveat. Deep Principle's comparison charts use GLM-5.3 rates to estimate Hydrogen costs. **Those estimates are not a verified Hydrogen-1.0 API price list.** The announcement does not establish the actual bill for a particular hosted or self-managed deployment.

## Quantized weights still need serious infrastructure

Deep Principle publishes separate NVIDIA NVFP4 and AMD MXFP4 checkpoints. These reduce precision for deployment, but their documentation describes different runtime configurations rather than one universal hardware requirement.

The [NVFP4 card](https://huggingface.co/deep-principle/Hydrogen-1.0-NVFP4) reports testing with SGLang on eight NVIDIA B300 GPUs. Its example deployment caps context at 262,144 tokens, below the model's stated one-million-token capacity. Raising that limit requires suitable runtime support and memory.

The [MXFP4 card](https://huggingface.co/deep-principle/Hydrogen-1.0-MXFP4) reports an AMD MI350X deployment using ATOM with tensor parallelism of four. It also documents a useful operational failure: 128 concurrent long-reasoning requests filled the KV cache pool and triggered eviction, alongside a sharp decline in answer quality. At 48 requests, utilization stayed below 75% and quality was normal in that test.

Those are observations from one configuration, not recommended concurrency limits for every deployment. They show why infrastructure validation belongs in the evaluation plan. A model that answers correctly at low load may behave differently when the serving system is under pressure.

## What researchers should evaluate next

The training approach preserves more of the research process, including plans, observations and revisions, rather than only final answers. Deep Principle uses supervised fine-tuning and reinforcement learning, with demonstrations involving external tools and particular research environments.

That suggests a practical evaluation question: does the model help a scientist make better next-step decisions on the lab's own held-out problems?

Useful evidence would include reproducible workflows, checked calculations, explicit assumptions and a record of where recommendations fail. A convincing result on one benchmark should lead to that investigation, rather than replace it.

Hydrogen's open weights make independent adaptation and evaluation possible. The next meaningful advance will be showing which scientific workflows retain their gains outside the team's own evaluation setup and under realistic serving conditions.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain research, model releases and engineering work in ways developers and potential customers can understand and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).