A 753B Science Model Shipped Under MIT. The Training Recipe Did Not.

Deep Principle, a Hangzhou company, released Hydrogen-1.0 on November 2. It is a 753 billion parameter mixture-of-experts model with about 40 billion parameters active per forward pass, built for chemistry, materials science and life sciences. The weights are on Hugging Face under the MIT license, behind a gate that approves access automatically.
That combination is unusual enough to be worth stating plainly. The largest open scientific model to date, under the most permissive common license, downloadable by anyone with eight GPUs and a request form.
It is also not what "open" usually means, and the gap between the two is the interesting part.
What was built and how
Hydrogen-1.0 starts from Z.ai's GLM-5.3, post-trained on domain data: chemistry, materials science, life sciences, plus records and agent trajectories from Deep Principle's own laboratories. The company runs an electronic lab notebook called Mira, and the model learned from full-cycle lab records captured through it, including failed experiments.
Training combined supervised fine-tuning with reinforcement learning, and the company says the infrastructure ran on both Nvidia CUDA and AMD ROCm. Megatron and SGLang are named as the training and serving frameworks.
The specification sheet is serious. Context window of one million tokens, maximum output of 128,000 tokens. BF16 serving tested on eight GPUs via SGLang. Quantized builds target specific accelerators: NVFP4 for B200 and B300, MXFP4 for MI350X and MI355X. The server exposes an OpenAI-compatible API with a reasoning effort setting.
That hardware profile puts self-hosting in the range of a research institute or a company with a dedicated cluster. Eight GPUs for BF16 inference is a departmental purchase, not a personal one. Which is why the other distribution channels matter: ModelScope hosts quantized variants, NeuroSpark offers an API, and the model also sits inside Deep Principle's own AI Scientist product, Mira.
The benchmark claims, with the correct labels
On FrontierScience-Research, a suite aimed at research-grade science problems rather than textbook questions, Hydrogen-1.0 scores 38.3 percent. GPT-6-Astra scores 33.3 percent. On GPQA-Diamond, a graduate-level science reasoning benchmark, a founder reports 93.0 percent, roughly matching Kimi-K3.
On three internal lab scenarios, the company says it beats GPT-6-Astra at maximum thinking by 5.13 points on scientific computing and 4.76 points on forward reaction prediction. It also publishes a loss: Astra at maximum thinking is 3.33 points better at NMR structure elucidation.
The cost figure is the one with the sharpest edge. About 25 cents per problem, against 88 cents for Astra Max. The blog plots accuracy against estimated API cost, which is a more honest chart than a bare leaderboard because it shows that the top of the frontier is expensive per unit of correctness.
Now the caveats, which are substantial and which the company states rather than hides.
Every number is self-reported. None has been reproduced by an outside group. Several come from internal question sets whose composition is not published. On the 117 experiment-planning questions used as a comparison, the only model included was Kimi-K3, because GPT-6-Astra and Opus 5.5 refused more than half the requests in that evaluation. That exclusion is defensible, since a model that will not answer cannot be scored, and it also means the headline comparison does not hold on that task.
The company also says differences across tasks should not be read as direct comparisons of model capability, which is the sort of sentence that deserves to be quoted more often.
Open weights, closed recipe
The model card carries one line that redefines the whole release: full datasets and recipes are not released with the weights.
So what is actually open? The parameters, in BF16 and two quantized formats. The license, which permits commercial use and derivative models with no royalty. The serving path, via an open framework. The benchmark claims, such as they are.
What is closed is the part that produced the capability. The data mixture. The reinforcement learning environment. The reward design. The curation of four years of laboratory records into training material. Anyone can run Hydrogen-1.0. Nobody can rebuild it.
This is now the dominant pattern in open model releases, and it is worth being clear about what it buys and what it costs. It buys reproducibility of results. A lab can download the weights, run the same problem, and check the claim, which is real progress over an API that changes underneath you. It buys sovereignty. A pharmaceutical company with sensitive data can run inference on its own hardware. It buys derivative work, since fine-tuning is permitted.
It does not buy science. If post-training on curated expert trajectories is where the capability comes from, then an open checkpoint is a finished artifact rather than a recipe, and the field's ability to improve on it is limited to the data it can produce itself.
The counterargument is that Deep Principle's data is not transferable anyway. A competitor cannot use another company's electronic lab notebook. In that framing, releasing weights is the maximum that can be released without giving away the only asset that matters.
Both readings are correct. The question is which one a given buyer needs.
The division of labor nobody planned
There is a structural detail here that will look obvious in two years. A specialist company in Hangzhou took a Chinese open base model from Z.ai, post-trained it on proprietary lab data, and shipped it under MIT with quantization builds for both Nvidia and AMD hardware.
Three years ago that pipeline did not exist. Foundation models came from a handful of labs, and everyone else built applications. Now the base layer is a commodity that specialists post-train on, and the specialization is the product. Hydrogen-1.0 is one of the clearest examples yet, and the fact that the base was Chinese while the distribution infrastructure is American and the license is MIT makes the supply chain genuinely multinational.
That has a practical consequence for buyers. When evaluating a domain model, the base model tells you less than it used to. What matters is the post-training data and whether the evaluation reflects your tasks. A 753B model that wins on FrontierScience-Research and loses on NMR structure elucidation is telling you exactly where its training data was strong.
How to decide
If you run a computational chemistry group and you have cluster access, download the weights and run ten of your own problems. Not the demo ones. The ones where you already know the answer and the ones where you do not.
The demo Deep Principle showed was a polyethylene melt simulation: the model wrote and executed the simulation code, then rendered a 3D molecular visualization and computed density, radial distribution function, radius of gyration and end-to-end distance. That is a competent junior computational scientist workflow. It is also a workflow a well-prompted frontier model can do.
The difference is what happens when you need to run it four hundred times on data you cannot send anywhere. That is the case Hydrogen-1.0 is built for, and it is a large enough case to justify the eight GPUs.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.