Deep Principle ships 753B science model Hydrogen-1.0 under MIT

Deep Principle has released a 753-billion-parameter scientific research model under an MIT license. Hydrogen-1.0 is a mixture-of-experts build on Z.AI's GLM-5.3, posted to Hugging Face with three weight packages.

The model card describes it as "a scientific problem-solving and research assistance model developed by Deep Principle based on GLM-5.3. Through text interaction, it is intended to assist with scientific questions, explain derivations, and organize responses to open-ended research problems."

One package is a BF16 unquantized release. A second, NVFP4, targets NVIDIA's Blackwell B200 and B300 accelerators. A third, MXFP4, targets AMD's Instinct MI350X and MI355X. The reference deployment runs SGLang across 8 GPUs with tensor parallelism. Context length is listed at 1M tokens, with a 128K-token maximum output.

The intended scope sits in a "Research in practice" section: "The technical blog presents examples of scientific computing, reaction-network exploration, NMR structure analysis, and experimental planning. These demonstrations combine the model with external tools and specific research environments; outputs require scientific review."

No benchmark names or scores appear on the page. The evaluation section says only that assessment "focuses on scientific multiple-choice questions and open-ended scientific research problems." On limits, Deep Principle writes: "The model may produce incorrect facts, calculations, equations, references, or scientific explanations. A detailed or confident response does not establish correctness."

This is the latest in a run of GLM-5.3 derivatives moving through our open-source AI tracking in recent weeks.