Deep Principle Hydrogen-1.0: An Open-Weight Science Model Built on GLM-5.3

Deep Principle has released Hydrogen-1.0, an open-weight model for chemistry, materials science and life sciences. In the company's launch post, it is a post-trained version of Z.ai's GLM-5.3, tuned for scientific discovery. The headline claim: it beats OpenAI's GPT-6-Astra on the FrontierScience-Research benchmark, at GLM-5.3 prices. Most of the other wins come from Deep Principle's own internal tests, so this post separates what the company claims from what you can check yourself.

If you build with open models, the release matters for one reason. It is a large, domain-tuned model that you can download, not just rent. If you follow AI for science, it is another sign that labs now compete on specialist post-training of strong open bases, not only on new pretraining runs.

TL;DR: what readers ask first

What Deep Principle actually released

Deep Principle describes Hydrogen-1.0 as "a start towards superintelligence for scientific discovery." The model is aimed at three fields: chemistry, materials science and life sciences.

The training recipe, as the blog describes it, has two stages. First comes supervised fine-tuning. Then comes reinforcement learning. The data mix is domain data, laboratory records from the company's own labs, and agent discovery trajectories, meaning logs of an AI agent working through a discovery task step by step. The post does not give data volumes.

The base model is GLM-5.3, the Z.ai model we covered when it tied Kimi K3 on an intelligence index. Deep Principle post-trains the whole mixture-of-experts network, not just an adapter layer. That matters, because it means the weights you download are a full model and not a small patch you must apply to GLM-5.3 yourself.

Inference weights come as BF16 and as 128 by 128 blockwise FP8, according to the blog. The Hugging Face organization also lists an NVFP4 build and an MXFP4 build for AMD hardware.

The benchmark claims, one by one

The company splits its results into public benchmarks and internal evaluations. The difference matters.

Public benchmarks

• FrontierScience-Research (FSR): Hydrogen-1.0 outperforms GPT-6-Astra, per the company. The text of the post gives no exact score in prose. It plots results against cost in its first figure.
• GPQA Diamond: Hydrogen-1.0 improves on GLM-5.3. The blog also says it roughly matches the open model Kimi-K3, without making clear which benchmark that covers.

Both claims are the company's reading of its own chart. We could not find an independent run of FSR on Hydrogen-1.0 yet. Treat the FSR win as promising, not settled.

Internal evaluations

Molecular cluster in a mortar standing for chemistry reaction prediction tasks in the Hydrogen-1.0 evaluation

Three details in the fine print change how you should read this table.

1. The company admits cross-task differences are not capability comparisons. Score gaps reflect how hard its internal cases are.
2. Two rivals were left out of the planning test. GPT-6-Astra and Opus-5.5 were omitted from the experiment-design comparison because they rejected more than 50% of requests. That is a real finding about refusals in chemistry, but it also means the planning result compares Hydrogen-1.0 only with Kimi-K3.
3. The cost figures exclude input and judge costs. The post says Hydrogen-1.0 uses GLM-5.3 pricing.

The honest summary: Hydrogen-1.0 looks strong in the domain it was trained for, and the company is open about one loss (NMR). But the evidence is vendor-reported. A reader should wait for replication on public sets before moving a workflow.

Claimed versus verified

On the license point, check the card of the exact variant you download. Different builds of the same model sometimes carry different license text, and the base GLM-5.3 license may flow through to some of them. Read it before shipping a product.

How to run it

The NVFP4 model card gives a concrete route. According to it:

• Format: NVFP4 for routed experts, FP8 for the dense MLPs in layers 0 to 2, and BF16 for shared experts, attention, router, embeddings, lm_head and the multi-token-prediction layer. The KV cache uses FP8. The team made it with NVIDIA ModelOpt.
• Hardware: a compatible NVIDIA Blackwell runtime. The team tested 8x B300. A separate MXFP4 build targets AMD Instinct and was tested on MI350X.
• Serving: an SGLang launch command with tensor parallel 8, served under the name hydrogen-1.0, plus an OpenAI-compatible client example.
• Context: the model is listed at 1M tokens with 128K max output, but the sample launch command caps deployed context at 262,144 tokens. If you need the full window, test it. The card itself warns that long context does not guarantee reliable use of every input detail.
• Hosting: the card says no inference provider deploys the model yet.

For most readers, the practical path is the hosted one: NeuroSpark for API calls, or Mira for the AI Scientist product. Both are named in the blog. Pricing for the API is stated only as following GLM-5.3 rates, so look up the current rate card.

A minimal client call, assuming you already run the SGLang server on your own machine:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8001/v1", api_key="none")
resp = client.chat.completions.create(
    model="hydrogen-1.0",
    messages=[{"role": "user", "content": "Propose a synthesis route for ..."}],
    max_tokens=4096,
)
print(resp.choices[0].message.content)

The model name and the OpenAI-compatible interface come from the card. The max_tokens value is illustrative.

Why a specialist post-train of an open base matters

Hydrogen-1.0 follows a pattern we have seen all year: a lab takes a strong open model and turns it into a domain specialist with its own data. We tracked the same logic in Gartner's claim that open models will run most agent steps, and in the AWS result where a 4B student matched a GLM-5.3 teacher on a shopping task.

The science angle is different in one way. The most valuable training signal here is proprietary: records from Deep Principle's own labs and its agents' discovery runs. Closed labs have the same advantage in principle. The frontier labs are building their own versions, such as OpenAI's GPT-Rosalind for life sciences, which ships behind trusted-access controls. Hydrogen-1.0 takes the other route: open weights and few access gates.

A ribbon folded into a compact shape, a picture of protein structure work that science models target

That choice has a dual-use side. A model that designs transition metal complexes and plans experiments can help a chemist. The refusal note in the blog shows the tension: Astra and Opus-5.5 rejected more than half of the planning prompts, while Hydrogen-1.0 answered. Open weights cannot be recalled. Deep Principle's model card says outputs are supporting material for review and not a basis for health or safety decisions, and that the lab-code path runs in a virtual environment before any physical execution. Those are sensible safeguards, but they are the company's own, and nobody else enforces them once the weights are public.

What this means for what you build or pay

• If you run chemistry or materials pipelines: try Hydrogen-1.0 on your own held-out cases next to your current model. The company's internal sets are not yours.
• If you pay for frontier APIs for science work: a refusal rate above 50% on planning prompts, as the blog reports for two closed models, is a cost too. Measure refusals on your own prompts.
• If you self-host: budget for 8 Blackwell-class GPUs for the NVFP4 build. This is not a laptop model.
• If you track AI for science: compare this release with the AlphaProof Nexus results on Erdos problems and with Anthropic's AI for Science expansion. Those are closed-lab efforts with credits and partnerships. Hydrogen-1.0 is an open-weight counterpart.
• If you evaluate models: note that FSR is the only public benchmark in the headline win. Wait for independent leaderboards before treating it as settled. For a sibling case of weights being edited by a frontier model, see GPT-6 Astra editing Qwen3.5-0.8B weights.

What is still open

• No independent FSR or GPQA Diamond reproduction yet.
• No parameter count in the blog. The 381B figure applies to the NVFP4 checkpoint on the card.
• No stated release date or license in the blog itself.
• No data volumes for the SFT and RL stages.
• No public detail on how the agent trajectories were filtered.

We will update this post when third-party numbers appear.

Related reading

• Z.ai GLM-5.3 ties Kimi K3 on an intelligence index
• Gartner and Revolut on open-weight models: claimed vs verified
• AWS SageMaker distillation: 4B model vs GLM-5.3 teacher
• OpenAI GPT-Rosalind system card addendum
• AlphaProof Nexus and Erdos problems
• Anthropic AI for Science expansion
• Official: Deep Principle launch post and the Hydrogen-1.0 NVFP4 model card

Model specs, licenses and prices are accurate as of November 4, 2026 and may change.