Hydrogen-1.0 explained: Deep Principle’s science model.

Hydrogen-1.0 is Deep Principle’s first open-weight model for scientific research. The company says it was built by post-training Z.ai’s GLM-5.3 on chemistry, materials science and life sciences data, plus records and agent trajectories from its own laboratories. Weights landed on November 2, 2026 under an MIT license, with a 1-million-token context window and quantized builds for NVIDIA Blackwell and AMD Instinct. Here is what Hydrogen-1.0 is, what Deep Principle claims is new, how to try it, and how it sits next to models already on the Atlas like GPT-6 Astra and Kimi K3.

What is Hydrogen-1.0?

Deep Principle describes Hydrogen-1.0 as a scientific problem-solving and research-assistance model. It is not trained from scratch: the Hugging Face model card lists GLM-5.3 as the base, and the card reports a model size of 753B parameters. The BF16 release is gated on Hugging Face (you agree to share account information to download) and mirrored on ModelScope. Deep Principle says full training datasets and recipes are not released with the weights.

According to the Hugging Face card, context length is 1M tokens and maximum output length is 128K tokens. Deployment was tested with SGLang on 8 GPUs. Separate NVFP4 and MXFP4 quantized repositories target NVIDIA Blackwell (B200, B300) and AMD Instinct (MI350X, MI355X) hardware.

What’s new in Hydrogen-1.0

Scientific post-training on lab work

Deep Principle’s blog says training combines supervised fine-tuning and reinforcement learning on four kinds of material: domain knowledge in chemistry, materials and life sciences; task-specific feedback from its scientists; full-cycle electronic laboratory notebook records (intent, procedure, observations and outcomes, including failures); and discovery trajectories that link choices to results over a campaign. The pitch is that open general models rarely see this kind of lab context, so post-training on it should improve experiment planning and research decisions.

Open weights for science

The company contrasts Hydrogen-1.0 with recent proprietary science demos (it cites OpenAI’s Navier–Stokes work and Anthropic’s enzyme discovery as examples of closed systems). Its stated goal is an open-weight model researchers can inspect, adapt and run themselves. The BF16 weights are under the MIT License on Hugging Face; Deep Principle notes independently authored code and docs under Apache 2.0.

Materials discovery loop

One showcase in the blog is transition metal complex (TMC) design. Each model starts from the same population and proposes new complexes over five rounds, seeing only the results of its own suggestions. Deep Principle reports that by round five the mean best-so-far HOMO–LUMO gap reaches 4.318 eV for Hydrogen-1.0, versus 3.689 eV for GPT-6 Astra and 3.353 eV for Kimi K3. These are Deep Principle’s own figures on an internal setup.

Hydrogen-1.0 benchmarks (Deep Principle’s numbers)

Every score below is self-reported by Deep Principle. Independent third-party runs were not available at launch. Treat them as the maker’s claims, not settled rankings.

• FrontierScience-Research: Deep Principle’s Chenru Duan reports 38.3% for Hydrogen-1.0 versus 33.3% for GPT-6 Astra.
• GPQA Diamond: Duan reports 93.0%; the blog says this roughly matches Kimi K3 and improves over GLM-5.3.
• Internal scientific computing (max thinking): Duan reports 76.9% for Hydrogen-1.0 versus 71.8% for GPT-6 Astra (a 5.13-point lead matching the blog).
• Internal one-step forward reaction prediction: 90.5% versus 85.7% for Astra at max thinking (blog: 4.76-point lead).
• Internal NMR structure elucidation: Astra ahead, 30.0% versus 26.7% for Hydrogen-1.0 (blog: 3.33-point gap).
• Experiment planning: the blog compares Hydrogen-1.0 with Kimi K3 on 117 shared questions; GPT-6 Astra and Claude Opus 5.5 were omitted because, Deep Principle says, they rejected more than 50% of requests in that evaluation.

The blog’s cost charts price Hydrogen-1.0 API usage using GLM-5.3 rates for comparison; Deep Principle has not published a separate public Hydrogen-1.0 token price in the announcement.

How to try Hydrogen-1.0

1. Download open weights from Hugging Face↗ (opens in a new tab) (gated) or ModelScope; keep generation_config.json so end-of-turn tokens work.
2. For local inference, Deep Principle documents an SGLang launch on 8 GPUs with tensor parallel 8, serving the model as hydrogen-1.0.
3. Use the NeuroSpark API if you want hosted access without running the weights yourself.
4. Try it inside Mira, Deep Principle’s AI Scientist product; Mira’s account has said promotional usage rates as low as 0.1× run from November 2 through November 30, 2026.
5. Quantized NVFP4 or MXFP4 builds are available if you are on Blackwell or Instinct hardware listed on the model card.

Who is Hydrogen-1.0 for?

Based on Deep Principle’s examples, it fits computational chemists and materials scientists who want an open model that can draft simulation workflows (the blog shows a polyethylene melt in GROMACS with topology repair), explore reaction networks, read NMR candidates, or plan lab protocols before a virtual “lab compilation” check. The model card stresses that outputs need scientific review: they are not verified experimental findings and should not drive health or safety decisions on their own.

What people are building with Hydrogen-1.0

At the time of writing there are no creator builds tagged to Hydrogen-1.0 on the Atlas yet—the model was added to the directory right after launch. We will list builds on that page as creators share them. For now, compare the scientific and agentic work people already ship with GPT-6 Astra and Kimi K3, two of the models Deep Principle uses as reference points.