Deep Principle Releases Hydrogen-1.0, a 753B Open-Weight Science Model
Deep Principle has released Hydrogen-1.0, a 753 billion parameter open-weight model for chemistry, materials science and life sciences. It is fine-tuned from Z.ai's GLM-5.3. The weights are on Hugging Face and ModelScope under the MIT License.
The release matters because most recent AI science results come from proprietary models. Deep Principle says Hydrogen-1.0 is an attempt at a reproducible, inspectable alternative. It is aimed at question framing, experiment planning, results analysis and open-ended discovery.
In one line: Deep Principle has released Hydrogen-1.0, an MIT-licensed 753 billion parameter open-weight scientific model post-trained from GLM-5.3, and it reports 38.3 percent on FrontierScience-Research, above GPT-6-Astra's 33.3 percent.
What is in the release
Hydrogen-1.0 is a mixture-of-experts model with 753 billion total parameters and about 40 billion active parameters, inherited from the GLM-5.3 base. It has a 1 million token context window and a 128,000 token maximum output, according to the model card. The weights ship under the MIT License, which means teams can download, run and modify the model without licensing restrictions.
Deep Principle calls Hydrogen-1.0 a scientific research assistant model. It is not a general chatbot. The stated targets are question framing, experiment planning, results analysis and open-ended discovery. It is not the first open model aimed at life science. For a smaller DNA-focused release, see Hugging Face's Carbon-X, a 14B open DNA language model. For teams that want lighter research utilities, the research and search tools directory has smaller options.
Training data and method
The training pipeline combines supervised fine-tuning with reinforcement learning. MarkTechPost lists four data sources. Domain knowledge covers chemistry, materials science and life sciences. Research scenarios come from task-specific feedback from Deep Principle scientists. Laboratory records are full-cycle electronic lab notebook records captured through Mira, Deep Principle's lab tool. Discovery trajectories include sequences of choices and feedback, including failed attempts.
The infrastructure runs on Megatron and SGLang across NVIDIA CUDA and AMD ROCm. The team keeps dynamic sparse attention causal across context-parallel shards. Muon optimizer state moves between CPU, GPU and disk depending on the backend. Reinforcement learning uses resumable rollouts, and each token keeps its original behaviour log-probability for importance correction. A layout adapter converts training checkpoints into BF16 and 128 by 128 blockwise FP8 inference weights.
One important gap: full datasets and recipes are not released. You get the weights, not the complete training data or exact recipe. That is still useful for inspection and fine-tuning, but it limits full reproducibility.
Benchmarks and real lab demos
Deep Principle reports that Hydrogen-1.0 scores 38.3 percent on FrontierScience-Research, compared with 33.3 percent for GPT-6-Astra. These are self-reported numbers. On GPQA-Diamond, a founder's tweet thread reports 93.0 percent. That improves on GLM-5.3 and roughly matches Kimi K3, per the blog. The comparison table also notes OpenAI reported 96.0 percent for GPT-6-Astra, so the ranking depends on which benchmark and which source you trust.
Internal research-task evals compare Hydrogen-1.0 with GPT-6-Astra at maximum thinking. Hydrogen-1.0 leads scientific computing by 5.13 percentage points and forward reaction prediction by 4.76 percentage points. GPT-6-Astra leads NMR structure elucidation by 3.33 points. In a transition metal complex design task, models chased a larger HOMO-LUMO gap over five rounds. By round five, 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 technical post shows four lab workflow demos. Hydrogen-1.0 writes Python to build a polyethylene melt and runs GROMACS. It detects a topology error, repairs it, and analyses the corrected 40 nanosecond trajectory. It grows an Ireland-Claisen reaction network of 27 structures through 30 plausible reactions. It ranks 7 candidate monoformylation products of 2-bromonaphthalene from a 1H NMR trace. It also writes a JSON protocol for a 12-vial Suzuki coupling study, and a virtual lab compilation step catches an aspiration without a fresh tip before execution. On 117 experiment-planning questions, Hydrogen-1.0 scored above Kimi K3. GPT-6-Astra and Opus-5.5 were excluded after refusing more than 50 percent of requests.
Deployment and hardware
Hydrogen-1.0 is not a laptop model. Deep Principle tested BF16 serving with SGLang on 8 GPUs using tensor parallelism of 8. NVFP4 builds target NVIDIA B200 and B300. MXFP4 builds target AMD Instinct MI350X and MI355X. The server exposes an OpenAI-compatible API with a reasoning_effort setting.
This is a serious infrastructure requirement for a small team. If you are running client delivery on a modest GPU budget, you will probably rent rather than self-host. The open weights still matter because you can inspect, fine-tune or serve wherever your data rules allow. Open-weight large models are becoming more common; I covered Mistral's trillion-parameter Large 4 earlier.
What it does not yet do
Hydrogen-1.0 still trails GPT-6-Astra on NMR structure elucidation by 3.33 points. Full datasets and recipes are not released. Benchmarks are self-reported, and the real lab demos come from Deep Principle's own technical post, not independent validation. The model is also too large for most local setups, and the 128,000 token output may need careful orchestration in long agentic runs.
For my work at Latent, I would not rush to deploy this. But I would watch two things: the open weights and whether an independent lab or research group reproduces the FrontierScience-Research claim. If a client in materials or pharma asks about an open scientific assistant, I would try Hydrogen-1.0 on a rented 8 GPU node with a small set of internal chemistry questions before touching any production workflow.
Quick answers
What is Hydrogen-1.0?
Hydrogen-1.0 is Deep Principle's open-weight scientific model with 753 billion total parameters and about 40 billion active parameters, fine-tuned from GLM-5.3 for chemistry, materials science and life sciences. It is MIT licensed.
How does Hydrogen-1.0 compare to GPT-6-Astra?
Deep Principle reports Hydrogen-1.0 scores 38.3 percent on FrontierScience-Research, above GPT-6-Astra's 33.3 percent. It still trails GPT-6-Astra by 3.33 points on NMR structure elucidation.
Can I run Hydrogen-1.0 locally?
Deep Principle tested BF16 serving with SGLang on 8 GPUs using tensor parallelism 8. NVFP4 builds target NVIDIA B200 and B300; MXFP4 targets AMD MI350X and MI355X. It is not a small workstation model.