What was actually released
Strip the benchmark chart away and the release itself is solid. Deep Principle, a Hangzhou-based company working on AI for materials research, took GLM-5.3, Z.ai's open model, and post-trained it for chemistry, materials science and life sciences. The weights are published on Hugging Face and ModelScope under the MIT License. The Hugging Face download is gated: users must accept the repository conditions and agree to share their account information. The weights can be inspected and fine-tuned, but access is not entirely anonymous or frictionless.
The model is large even by open-weight standards. Hugging Face lists 753 billion parameters, arranged in a mixture-of-experts architecture. Release coverage puts the active count at roughly 40 billion parameters per token, consistent with the GLM-5.3 base. The model card describes deployment tested with SGLang across eight GPUs, a headline context length of 1 million tokens and output up to 128,000 tokens.
The 1-million-token figure is the model's listed context capacity, not a guarantee that every deployment can serve that much context. The model card says actual budgets depend on the inference framework, server configuration and available resources. Its tested SGLang setup uses eight GPUs; laboratories should not treat the headline context length as a turnkey serving specification.
Quantized builds narrow the hardware choice rather than widening it: NVFP4 for NVIDIA B200 and B300, MXFP4 for AMD Instinct MI350X and MI355X. For labs without a cluster, the realistic doors are the NeuroSpark API, whose price has not been disclosed, and Mira, Deep Principle's own AI Scientist product.
Every score is the maker's
The headline claim is that Hydrogen-1.0 beats GPT-6-Astra on FrontierScience-Research, 38.3% to 33.3%. Those exact figures are plotted in Figure 1 of Deep Principle's technical blog; the accompanying text states the win but does not print the scores. Both measurements come from Deep Principle's evaluation, not an independent head-to-head reproduction. The chart also compares estimated API costs, but Hydrogen-1.0 is priced using GLM-5.3's output-token rates because its own NeuroSpark API price has not been disclosed. Input and judge costs are excluded.
The rest of the scoreboard is more mixed than the headline. On GPQA-Diamond, Deep Principle's blog claims improvement over GLM-5.3 and rough parity with Kimi-K3. A 93.0% Hydrogen-1.0 score was reported by the company's founder, while OpenAI has reported 96.0% for GPT-6-Astra. These are separately reported results, not a controlled head-to-head evaluation, and should not be treated as directly comparable without matching evaluation details. On the company's three internal lab scenarios, Hydrogen-1.0 leads Astra at maximum thinking on scientific computing and forward reaction prediction, and loses on NMR structure elucidation. The 117-question experiment-planning comparison is narrower still: Hydrogen-1.0 was compared with Kimi-K3 only. GPT-6-Astra and Opus-5.5 were excluded because they rejected more than 50% of requests in the evaluation. Deep Principle discloses that limitation, but the result cannot establish that Hydrogen-1.0 outperforms those omitted models at experiment planning.
None of this makes the claims false. It makes them unverified in a specific, fixable way: the weights are public, so anyone with eight suitable GPUs can rerun every test. Until someone does, the correct label for 38.3% is "Deep Principle reports", not "Hydrogen-1.0 scores".
The training data is the real claim
The more durable part of the release is not a percentage. Hydrogen-1.0 was trained on four things: domain knowledge, research scenarios with feedback from Deep Principle's scientists, full-cycle electronic lab notebook records, and discovery trajectories, the sequence of choices, results and revisions behind a research campaign, failed attempts included. Papers record what worked. Lab notebooks record what was tried, in what order, and what happened next. That is the material Deep Principle believes can help a model move beyond recalling published results toward planning experiments: not just what worked, but what was attempted, what failed and what the researchers tried next. Whether those records are responsible for the reported performance gains remains an open question, because the company has not released the complete training data or recipes needed to isolate their contribution.
The demonstrations are built to show exactly that. The model writes Python to build a polyethylene melt, runs GROMACS, catches a topology error, repairs it and analyzes a corrected 40-nanosecond trajectory. It drafts a JSON protocol for a 12-vial Suzuki coupling study, and a virtual rehearsal catches it trying to aspirate without a fresh tip before anything reaches a real bench. These are company-selected examples, and the card is blunt about the limits: outputs may contain incorrect facts, calculations and references, and should not be treated as verified experimental findings.
The openness has a boundary, and the card states it plainly: the full datasets and complete training recipes are not released with the weights. The release is inspectable at the level of the model and closed at the level of the method that produced it. The weights are open; the training process is not fully reproducible. A rival lab can inspect the released model, fine-tune it and independently evaluate its behavior. It cannot reproduce the original training run from the published material because the complete datasets and training recipes remain private.
What would settle it
One independent run. FrontierScience-Research is a public benchmark, the weights are downloadable under MIT, and the serving stack is documented. If an outside group reproduces anything near 38.3%, Deep Principle has done something real: taken an open base model and, with lab records rather than scale, moved it past a frontier proprietary system. If the number collapses under outside grading, the release becomes a case study in how much a chart can carry when the maker runs the test and prices its model at another model's rates.
Either outcome is useful, which is the strongest thing that can be said for an open release. Closed models ask to be believed. This one can be checked, and now needs to be.
Sources
- [1] Deep Principle, "Introducing Hydrogen-1.0", company technical blog, November 2, 2026 (read first-hand in full before publication — post-training of GLM-5.3 on domain data, laboratory records and agent trajectories; the FrontierScience-Research win stated in words with scores in Figure 1; internal-task margins of +5.13 and +4.76 points and the −3.33 NMR result; the 117-question planning comparison and the refusal-based omission of GPT-6-Astra and Opus-5.5; the 4.318 / 3.689 / 3.353 eV HOMO-LUMO results; the GROMACS, reaction-network, NMR and Suzuki-protocol demonstrations; the Figure 1/2 cost methodology using GLM-5.3 pricing; the four access routes; UniPat, Mind Lab and NeuroSpark credits)Read source
- [2] Deep Principle, Hydrogen-1.0 model card, Hugging Face (read first-hand in full before publication — GLM-5.3 base, 1M-token context, 128K maximum output, MIT License for the weights and Apache 2.0 for Deep Principle code and documentation, BF16 / NVFP4 / MXFP4 versions and their hardware, SGLang deployment tested on 8 GPUs, the statement that full datasets and recipes are not released, and the warnings that outputs may be incorrect and should not be treated as verified experimental findings)Read source
- [3] MarkTechPost, "Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery", November 2, 2026 (read first-hand in full before publication — 753B parameters with roughly 40B active per the GLM-5.3 base; the 38.3% vs 33.3% FrontierScience-Research figures as self-reported; the 93.0% GPQA-Diamond figure from a founder's thread; the comparison table listing GPT-6-Astra's GPQA at 96.0% as OpenAI-reported; training-data sources and infrastructure detail)Read source
- [4] TSN, "Hydrogen-1.0: Deep Principle's Open Science Model", November 3, 2026 (read first-hand in full before publication — the access-condition note for the Hugging Face download, the chart-based presentation of the benchmark claims, the internal-test margins and losses, the planning-test exclusion, and the model card's warning language)Read source
- [5] TokenStrom, "Hydrogen-1.0: 753B open-weight chemistry and materials model under MIT license", November 2026 (corroboration of the 753B / ~40B-active specification, the benchmark table including GPQA-Diamond, the HOMO-LUMO and experiment-planning details, the eight-GPU SGLang test and the undisclosed API price)Read source
- [6] NeuralBrew, "Deep Principle Releases Hydrogen-1.0, an Open-Weight Scientific Model Derived from GLM-5.3", November 2, 2026 (corroboration of the release, the lab-notebook training data, the 38.3% / 33.3% and 93.0% figures attributed to the developer and to Chenru Duan, the internal-task margins, and the four access channels)Read source