Deep Principle Opens Hydrogen-1.0, a 753B-Parameter Science Model

Most of the AI models credited with scientific results this year are closed. Deep Principle is betting the other way. Its Hydrogen-1.0 is a 753-billion-parameter model for chemistry, materials science and life sciences whose weights anyone can download from Hugging Face under the MIT licence, after agreeing to share their Hugging Face username and email address.

Hydrogen-1.0 is not built from scratch. The company’s technical blog says it post-trained GLM-5.3, an open model from Z.ai, on “domain data from chemistry, materials science, and life sciences, together with records and agent trajectories from our own laboratories”. Every performance figure it gives comes from Deep Principle’s own evaluations, several on internal question sets, and none has yet been reproduced by an outside group.

What Hydrogen-1.0 is and how to get it

The Hugging Face model card lists 753 billion parameters in BF16 format, with smaller NVFP4 and MXFP4 quantised versions in separate repositories. Context runs to 1 million tokens, with up to 128,000 tokens of output, although the card notes that usable budgets depend on hardware and settings. It was tested with the SGLang serving software on eight GPUs, which puts self-hosting out of reach for most individual researchers.

Besides Hugging Face, the model is available on ModelScope, through an API from NeuroSpark, and inside Deep Principle’s own AI Scientist product, Mira. The company says it worked with UniPat and Mind Lab on data, training pipeline and infrastructure, and with NeuroSpark on training infrastructure. Training combined supervised fine-tuning with reinforcement learning, and the blog says its training infrastructure supports both NVIDIA CUDA and AMD ROCm hardware.

The release is open in its weights, not in full. The model card says “full datasets and recipes are not released with the weights”.

Deep Principle’s benchmark claims

According to the company, Hydrogen-1.0 “outperforms GPT-6-Astra on FrontierScience-Research (FSR) and improves over GLM-5.3 on GPQA-Diamonds in our evaluations, roughly matching the open frontier model Kimi-K3”. The blog presents these results in a chart rather than in its text, and it plots scores against estimated API cost.

On three internal lab scenarios, Deep Principle says Hydrogen-1.0 beats GPT-6-Astra at maximum thinking by 5.13 percentage points on scientific computing and 4.76 points on forward reaction prediction. The company also reports a loss: “Astra at maximum thinking still scores higher on NMR structure elucidation by 3.33 points.” It adds that differences across tasks “should not be read as direct comparisons of model capability”.

For planning lab experiments, the company compared Hydrogen-1.0 only with Moonshot’s Kimi-K3, on 117 shared questions. GPT-6-Astra and Opus-5.5 were left out, it says, because they “rejected more than 50% of requests in this evaluation”. In a materials test, models proposed transition metal complexes over five rounds; Deep Principle says Hydrogen-1.0 reached a mean best-so-far HOMO–LUMO gap of 4.318 eV, against 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3.

The blog also shows Hydrogen-1.0 writing and repairing a polymer simulation in GROMACS, mapping a 30-reaction rearrangement network and writing a robot protocol for a 12-vial Suzuki coupling study that is checked in a virtual lab before it runs.

An open science model whose results still need outside tests

The release matters more for what it makes possible than for the scores it reports. The blog itself names the gap: the models behind headline AI science results are proprietary, and “establishing comparable, independently reproducible scientific discovery with general-purpose open-weight models remains an important aspect”. Open weights let any lab rerun Deep Principle’s tests, probe failures and fine-tune on its own data, which a closed API does not allow.

That is also why the company’s numbers should be read cautiously. Most of the comparisons use internal question sets that outsiders cannot see. The experiment-planning test drops the two models that refused most requests, so it compares Hydrogen-1.0 only with Kimi-K3. The materials result covers one design task over five rounds. None of this means the claims are wrong, but none of it has been checked by an outside group. Moonshot’s own tests for Kimi K3.1 Pro and Mistral’s 1-trillion-parameter Large 4 show how often open releases now arrive with vendor-run benchmarks first and independent ones later.

The approach is a familiar one for Z.ai’s model family, which already has a legal specialist built on GLM-5.3-Flash. What is new is the training on laboratory records and full discovery trajectories, including failed attempts, rather than on papers alone. Closed labs are pushing the same idea through programmes such as Anthropic’s AI for Science credits and tools such as Xaira’s X-Scientist. The model card’s own warning is the right frame: outputs “should not be treated as verified experimental findings”.

The first outside benchmark runs will settle the claims

Deep Principle says its next step is to make what the model learns in one discovery campaign carry over to other targets and material families. For outside researchers, the immediate question is simpler: whether Hydrogen-1.0’s lead over GPT-6-Astra survives when someone other than Deep Principle runs the tests.

Sources

• Deep Principle: “Introducing Hydrogen-1.0” (company technical blog)
• Hugging Face: deep-principle/Hydrogen-1.0 model card (official model repository)

Related stories

• Z.ai Launches GLM-Legal, an Agentic Legal Model Built on GLM-5.3-Flash
• Moonshot’s Own Tests Put Kimi K3.1 Pro at 65.7% on Terminal-Bench
• Mistral Opens Large 4 Weights: a 1-Trillion-Parameter Model
• Anthropic Expands AI for Science Credits to 20 Research Partners