Deep Principle, a Hangzhou AI-for-materials company, released Hydrogen-1.0 on Hugging Face on November 2. It's GLM-5.3 from Z.ai, post-trained on domain knowledge, research scenarios, laboratory records and discovery trajectories. The model card lists 753B parameters, a 1M-token context window and 128K tokens of maximum output. The weights are MIT licensed; the training data and full recipes are not released.
There are three versions: BF16 weights, an NVFP4 build for NVIDIA B200 and B300, and an MXFP4 build for AMD MI350X and MI355X. Deep Principle says it tested serving with SGLang on 8 GPUs at tensor parallelism 8, and the server exposes an OpenAI-compatible API with a reasoning_effort setting. The card has one practical warning: keep generation_config.json, because it holds the end-of-turn tokens and without it the model doesn't stop at the end of an answer.
On scores, MarkTechPost's summary of Deep Principle's results puts Hydrogen-1.0 at 38.3% on FrontierScience-Research against 33.3% for GPT-6 Astra, roughly level with Kimi-K3 on GPQA-Diamond, and 3.33 points behind GPT-6 Astra (max thinking) on NMR structure work. All of those are self-reported. The card itself says outputs need scientific review and shouldn't be treated as verified findings.