Deep Principle has released Hydrogen-1.0, a 753B parameter open-weight model designed for chemistry, materials science, and life sciences. The system is built by post-training Z.ai’s GLM-5.3 base on domain-specific data, laboratory records, and agent discovery trajectories. Weights are available under the MIT License.
Key specifications and performance
The model operates with approximately 40B active parameters per request. It supports a 1M-token context window and can generate up to 128K tokens. Testing on SGLang with 8 GPUs indicates support for NVFP4 on NVIDIA B200/B300 hardware and MXFP4 on AMD Instinct MI350X/MI355X chips.
Deep Principle reports a FrontierScience-Research score of 38.3% for Hydrogen-1.0. This beats GPT-6-Astra, which scored 33.3% in the same test. On the GPQA-Diamond benchmark, the model achieved 93.0%, matching the performance of Kimi-K3.
In internal evaluations comparing maximum thinking modes, Hydrogen-1.0 led GPT-6-Astra by 5.13 percentage points in scientific computing tasks and by 4.76 points in forward reaction prediction. However, GPT-6-Astra outperformed Hydrogen-1.0 by 3.33 points on NMR structure elucidation.
In a transition metal complex design task, models were asked to chase a larger HOMO-LUMO gap over five rounds. By the fifth round, Hydrogen-1.0 reached a mean best-so-far gap of 4.318 eV. GPT-6-Astra reached 3.689 eV and Kimi-K3 reached 3.353 eV. The model also scored higher than Kimi-K3 on 117 experiment-planning questions. GPT-6-Astra and Opus-5.5 were excluded from some comparisons after refusing more than 50% of requests.
Training data and infrastructure
The development pipeline combines supervised fine-tuning with reinforcement learning. Data sources included domain knowledge from chemistry and life sciences, task-specific feedback from Deep Principle scientists, full-cycle electronic lab notebook records captured through Mira, and discovery trajectories showing sequences of choices and feedback, including failed attempts.
Infrastructure relies on Megatron and SGLang across NVIDIA CUDA and AMD ROCm. The team maintains dynamic sparse attention causal across context-parallel shards. The Muon optimizer state moves between CPU, GPU, and disk depending on the backend. Reinforcement learning uses resumable rollouts, and each token retains its original behavior log-probability for importance correction. A layout adapter converts training checkpoints into BF16 and 128×128 blockwise FP8 inference weights. Full datasets and recipes are not released.
Practical lab workflow examples
Demos show the model writing Python code to build a polyethylene melt and run GROMACS. It detects a topology error, repairs it, and analyzes the corrected 40 ns trajectory. The system grew an Ireland-Claisen reaction network of 27 structures through 30 plausible reactions. It ranked seven candidate monoformylation products of 2-bromonaphthalene from a 1H NMR trace. A virtual lab compilation step caught an aspiration without a fresh tip before execution.
Deployment and API details
Weights are hosted on Hugging Face and ModelScope. Deep Principle tested BF16 serving with SGLang on 8 GPUs using a tensor parallelism of 8. The server exposes an OpenAI-compatible API with a reasoning_effort setting.
Comparison with other models
While GLM-5.3 and Kimi-K3 are open-weight, GPT-6-Astra remains proprietary. Hydrogen-1.0 focuses specifically on chemistry, materials, and life sciences, whereas GLM-5.3 targets coding and agents, and Kimi-K3 is general-purpose with vision capabilities. Pricing for GLM-5.3 via Z.ai starts at $1.40 per million input tokens and $4.40 for output. Kimi-K3 charges $3.00 for input and $15.00 for output per million tokens. OpenAI charges $10.00 for input and $50.00 for output.
What it means
Researchers now have a fully open alternative to proprietary science models. The inclusion of actual lab notebooks and failed attempts in the training data addresses a gap where previous models relied only on published papers. This allows for better reproducibility and inspection of how scientific decisions are made. However, users must still provision significant GPU hardware to run the model locally, as no cloud deployment details are provided.
