A Chinese Lab Open-Sourced a 753-Billion-Parameter Science Model That Beats GPT-6 Astra on One Benchmark

Deep Principle, a Hangzhou startup building AI tools for materials and chemistry research, released Hydrogen-1.0 on Monday: a 753-billion-parameter model post-trained from Z.ai's GLM-5.3 and published under the MIT license. The weights are on Hugging Face and ModelScope in roughly 1.5 terabytes of BF16 shards numbering 303, alongside quantized NVFP4 and MXFP4 builds aimed at B300 and MI350X hardware. The model takes a one-million-token context and can emit up to 128,000 tokens.

The training recipe is the interesting part. Deep Principle says it combined domain literature in chemistry, materials science and the life sciences with task-level feedback from working scientists, its own electronic lab notebooks and full discovery trajectories — including experiments that failed and hypotheses that were revised. Most science models are trained on papers and textbooks, which contain tidy answers. Lab notebooks contain the sequence that actually produces a result: what was tried, what broke, what was changed next.

The published scores are strong but self-reported. On FrontierScience-Research, a benchmark of unfamiliar multi-step problems in physics, chemistry and biology, Hydrogen-1.0 reaches 38.3 percent against 33.3 percent for GPT-6 Astra and 24.4 percent for the unmodified GLM-5.3 base. On GPQA-Diamond, expert-written multiple-choice science questions, it scores 93.0 percent, about three points above its base and roughly level with Kimi-K3 at 93.5 percent. On Deep Principle's own three internal tasks it leads Astra by 5.13 points on scientific computing and 4.76 points on single-step reaction prediction, and trails by 3.33 points on NMR structure elucidation. The company notes that absolute scores across different internal tasks reflect task difficulty and should not be read as a like-for-like capability comparison.

The demos are more persuasive than the leaderboard. Given a polymer physics question, Hydrogen-1.0 wrote a Python pipeline to build a polyethylene melt, ran it through GROMACS, detected a topology error mid-run, diagnosed it, fixed the definition and restarted the simulation, then analysed density, packing, diffusion and chain dimensions over a 40-nanosecond trajectory. In another case it laid out an Ireland-Claisen rearrangement network of 27 molecular structures and 30 reaction records and compared competing routes. It also caught a procedural error while compiling a 12-tube Suzuki coupling protocol — a pipetting step that reused a tip — before anything ran on a bench.

The most concrete claim is a closed search loop. Each model started from the same population of transition-metal complexes, recommended one candidate per round, and received only its own results as feedback. By round five, Hydrogen-1.0's best candidate reached a HOMO-LUMO gap of 4.318 eV, against 3.689 eV for GPT-6 Astra and 3.353 eV for Kimi-K3, which the company presents as evidence that optimisation behaviour can be trained rather than prompted.

The engineering choices are unusually candid for a release note: a Muon optimiser with FP32 master weights and momentum held on CPU and shuttled to the GPU one matrix at a time, momentum offloaded to disk on ROCm to free memory for long-sequence activations, offline sequence packing without truncating reasoning traces, cross-rank causal constraints for dynamic sparse attention, and a two-way adapter that bridges BF16 training weights and FP8 inference weights in 128-by-128 blocks.

The licence gap is worth flagging. GLM-5.3 ships under Z.ai's own terms, which now include a compliance review regime for model-as-a-service use; its derivative here is MIT, which allows commercial use, modification and redistribution. That is the most permissive licensing in this class of large open-weight science model, and it shifts the moat from the base model to private process data — which Deep Principle has not released. Independent replication is still missing entirely, and the NMR result is a reminder that the model is not uniformly better.

Pricing matches GLM-5.3, and the model is reachable through Hugging Face, ModelScope, the NeuroSpark API and Deep Principle's own AI scientist product, Mira. The larger signal is directional: the discovery-loop capability that frontier labs have demonstrated behind closed APIs — agents that design experiments, debug instrument chains and iterate on their own results — now has an open-weight version that any lab can download and continue training on its own data.