Is Hydrogen-1.0 state of the art?Pending benchmarks.Announced Nov 2, 2026; no independent results yet.
Deep Principle's open-weights model for scientific research in chemistry, materials science and life sciences, post-trained from Z.ai's GLM-5.3 on domain data, laboratory records and discovery trajectories. It has a 1M-token context and up to 128K output tokens. Weights on Hugging Face and ModelScope under the MIT license, with NVFP4 and MXFP4 builds; Deep Principle also offers it through NeuroSpark's API and its Mira AI Scientist product.
The details so far
- Lab
- Deep Principle
- Announced
- Nov 2, 2026
- API
- Live · NeuroSpark API (model hydrogen-1.0)unsure · linked from Deep Principle's blog; NeuroSpark's model page needs sign-in
- Open weights
- Published · Hugging Face deep-principle/Hydrogen-1.0 (BF16, gated with automatic approval) and ModelScope deep-principle/Hydrogen-1.0
- Parameters
- About 753B (753,329,940,480 parameters in the BF16 safetensors, the same count as the GLM-5.3 checkpoint on Hugging Face)
- Type
- Text reasoning model for scientific work: experiment planning, protocol writing, results analysis, simulation workflows and materials design
- Base model
- zai-org/GLM-5.3 (Z.ai), fully post-trained by Deep Principle; same architecture (Hugging Face tag glm_moe_dsa)
- Training
- Supervised fine-tuning and reinforcement learning on chemistry, materials science and life sciences data, feedback from Deep Principle scientists, laboratory records and discovery trajectories; datasets and recipes not released
- Context
- 1M tokens, up to 128K output tokens, per the Hugging Face card
- Variants
- deep-principle/Hydrogen-1.0-NVFP4 (for NVIDIA Blackwell B200 and B300) and deep-principle/Hydrogen-1.0-MXFP4 (for AMD MI350X and MI355X), not gated
- License
- MIT for the weights (LICENSE, copyright Deep Principle and Z.AI); Deep Principle's own code and docs under Apache 2.0; the upstream GLM-5.3 checkpoint keeps its own license
- Serving
- SGLang on 8 GPUs, served model name hydrogen-1.0, per the Hugging Face card
- Product
- Available in Mira, Deep Principle's AI Scientist
- Collaborators
- UniPat and Mind Lab (data, training pipeline and infrastructure) and NeuroSpark (training infrastructure), per Deep Principle
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 38.3% on FrontierScience-Research, against 33.3% for GPT-6-Astra, and 93.0% on GPQA-Diamond, per Deep Principle's CTO.
Lab claim: in Deep Principle's internal tests, 5.13 and 4.76 points ahead of GPT-6 Astra (maximum thinking) on scientific computing and forward reaction prediction, while Astra scores 3.33 points higher on NMR structure elucidation.
Lab claim: in transition metal complex design, a mean best-so-far HOMO–LUMO gap of 4.318 eV after five rounds, against 3.689 eV for GPT-6-Astra and 3.353 eV for Kimi-K3.
Lab claim: a higher mean rubric score than Kimi-K3 on 117 shared experiment-planning questions.
Sources
LABDeep Principle announcement on Xx.com/DeepPrinciple/status/2117249008972550447 · Nov 2, 2026LABChenru Duan (Deep Principle CTO) post on Xx.com/chenru_duan/status/2117249292469665838 · Nov 2, 2026LABIntroducing Hydrogen-1.0 (Deep Principle blog)blog.deepprinciple.com/introducing-hydrogen-1.0/REPOdeep-principle/Hydrogen-1.0 on Hugging Facehuggingface.co/deep-principle/Hydrogen-1.0REPOdeep-principle/Hydrogen-1.0-NVFP4 on Hugging Facehuggingface.co/deep-principle/Hydrogen-1.0-NVFP4REPOdeep-principle/Hydrogen-1.0-MXFP4 on Hugging Facehuggingface.co/deep-principle/Hydrogen-1.0-MXFP4REPOdeep-principle/Hydrogen-1.0 on ModelScopemodelscope.cn/models/deep-principle/Hydrogen-1.0PRESSModelScope post on Xx.com/ModelScope2022/status/2117164818880696417 · Nov 2, 2026Compare when it lands
Unscored: its base model, GLM 5.3. A scored comparison appears once both are on a shared board.
Tracking continues until an independent board lists it.