A free AI model for scientists comes with a catch for companies earning over $10 billion

Deep Principle, a Hangzhou startup building AI for scientific discovery, released Hydrogen-1.0 on November 2. It is a 753-billion-parameter model aimed at chemistry, materials science and the life sciences. It was not built from scratch: it is Z.ai's GLM-5.3 given further specialist training. The Hugging Face model card says its weights, the trained model files, are under the permissive MIT License.

Uptake has been fast for a model this size. On November 5, Hugging Face's API reported 19,960 downloads for the main repository, while the model page itself displayed a lower 11,798. Either way, that is a lot for a model whose card was tested on an eight-GPU server.

For companies, the "MIT" tag is both the main attraction and the part that needs the closest reading. The base model is not MIT-licensed. Every headline performance figure comes from Deep Principle's own evaluations. And the full-precision weights sit behind a sign-in wall. That does not make the release unimportant, but it is different from how the label reads.

What companies can actually get

Anyone can run the model, but no outsider can rebuild it. Deep Principle published full-precision weights plus compressed versions for Nvidia's Blackwell B200 and B300 and AMD's Instinct MI350X and MI355X chips. Training used four kinds of data: domain knowledge, task feedback from the company's own scientists, electronic lab-notebook records captured through its Mira platform, and complete discovery trajectories, failed attempts included.

That data is the company's moat, and it stays private. The card says the full datasets and training recipes are not released. Access is also less open than the license suggests. Downloading the main full-precision files requires a Hugging Face login and agreeing to share contact details with Deep Principle. For hosted use, the company points to an API through partner NeuroSpark and to its own Mira "AI Scientist" product. Hugging Face listed no third-party provider running the model as a service.

Every performance claim is the company's own

Buyers and investors should treat the performance story as unverified. Runtimewire and DiarioBitcoin both report the company's headline numbers: 38.3% on FrontierScience-Research against 33.3% for OpenAI's GPT-6-Astra, and 93% on GPQA-Diamond. The blog itself gives those figures only in a chart. It describes Hydrogen as roughly matching Moonshot's Kimi-K3 on the public benchmarks. DiarioBitcoin's comparison table lists 96% for GPT-6-Astra on GPQA-Diamond, attributed to OpenAI, so the science model does not lead there.

On internal tasks, Deep Principle says Hydrogen beats GPT-6-Astra at its maximum thinking setting by 5.13 percentage points in scientific computing and 4.76 in one-step forward reaction prediction. It trails by 3.33 on NMR structure elucidation. The company itself warns that absolute scores reflect how hard its internal cases are.

All of these are the company's own measurements, scored partly by an AI judge, and the Hugging Face card publishes none of them. No independent replication had appeared by publication. The two headline tests are public, so outside groups can check those claims, as Runtimewire also noted. The internal tests cannot be checked the same way.

One detail matters for anyone weighing safety risk. Deep Principle says it left GPT-6-Astra and Anthropic's Opus-5.5 out of its 117-question experiment-design comparison. In the blog's words, they "rejected more than 50% of requests" in that evaluation. There are two readings. Open science models may be useful exactly where closed models' safety filters are strictest. Or the questions reach into wet-lab territory that frontier labs consider risky. The blog does not publish the questions, so readers cannot tell which applies. Either way, MIT-licensed weights cannot be recalled once downloaded.

A permissive label on a stricter foundation

The license risk falls on the very largest companies. Z.ai released GLM-5.3's weights on August 28 under a custom "GLM-5.3 License." As Digital Applied documented, that broke with recent Z.ai practice: the smaller GLM-5.3-Flash had shipped two days earlier under unmodified MIT. The flagship license grants MIT-style rights, including modified and derivative versions, with one substantive condition.

It applies if the licensee or an affiliate runs a model-as-a-service business, giving third parties API-style access with meaningful control over inputs, parameters or training data. If combined revenue "exceeds 10 billion US dollars (or the equivalent in other currencies) in total over any consecutive 12 months," the licensee "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." The test is total revenue, not API revenue. The review's scope and method "shall be reasonably determined by Z.AI." Digital Applied found no published criteria, timeline or appeal process. Products that merely embed model features fall outside the definition, as do services that just relay requests to someone else's hosted model.

Hydrogen's public card does not mention Z.ai's review clause. Its separate licensing files could not be retrieved without signing in. Because the clause names derivative works explicitly, a plain reading suggests a company over the revenue line, running an API business and using Hydrogen commercially, would still need Z.ai's review, whatever the label says. That reading has not been tested, and neither company has addressed it publicly. For nearly everyone else, from university labs to startups, the threshold never applies and the practical terms are permissive.

The bigger problem is the labels compliance teams rely on. Hugging Face lists Hydrogen among 21 retrained versions of GLM-5.3, but its license field shows only the top layer. A compliance team filtering by license tag sees "MIT" and moves on. As more freely downloadable models are built on a few large base models, license terms can quietly change from one publisher to the next. Tooling and procurement checklists have not caught up.

What to watch

Four things will show whether Hydrogen holds up. Independent reruns of the two public benchmark claims will test whether specialized training beats general frontier models. A statement from Deep Principle, Z.ai, or both on how the GLM-5.3 clause applies to MIT-tagged derivatives would settle the license question. Hosting providers adding the model would show demand beyond curiosity downloads. And a sample of the experiment-design questions the closed models refused would let outsiders judge whether the model's edge comes from capability, from fewer safeguards, or from both.