Hydrogen-1.0 Explained: Deep Principle’s 753B Open-Weight Science Model vs GPT-6 Astra

Hydrogen-1.0 Explained: Deep Principle’s 753B Open-Weight Science Model vs GPT-6 Astra

On 2 November 2026, a small Hangzhou startup called Deep Principle released Hydrogen-1.0, a 753-billion-parameter artificial intelligence model built for one job only: scientific research. It is not a chatbot, it is not a coding assistant, and it will not write your emails. Instead, the model is trained to plan chemistry experiments, predict how molecules will react, read long research papers, and hunt for new materials with the properties a researcher actually wants. The weights are published under the permissive MIT licence, free for anyone to download and use commercially — with one very large catch. The model is so big that it was tested on eight high-end GPUs, and the download alone is roughly 452 GB.

This guide explains what the model is, how Deep Principle built it from Z.ai’s GLM-5.3, where it beats OpenAI’s GPT-6 Astra on the company’s own benchmarks, what it costs to run, and who should — and should not — try to use it.

Scientist holding a beaker in a chemistry lab, the kind of research Hydrogen-1.0 is built to support

Table of Contents

Hydrogen-1.0 at a Glance

If you only have sixty seconds, here is the whole story of Hydrogen-1.0 in one table. Everything below expands on these points, with the honest caveats attached.

Detail Hydrogen-1.0
Released 2 November 2026 (announced on X by founder and CTO Chenru Duan; press release followed on 4 November)
Developer Deep Principle, a materials-research AI company based in Hangzhou, China
What it is An open-weight large language model specialised in chemistry, materials science and life sciences
Base model Z.ai’s GLM-5.3, post-trained on scientific domain data and Deep Principle’s own laboratory records
Size 753 billion total parameters (mixture-of-experts), about 40 billion active per token
Context window 1 million tokens, with up to 128,000 output tokens (depends on framework and hardware)
Licence MIT licence for the weights; download is gated — you agree to share your Hugging Face account information
Where to get it Hugging Face and ModelScope (weights), NeuroSpark (API), Mira (Deep Principle’s AI Scientist product)
Headline benchmark 38.3% on FrontierScience-Research vs 33.3% for GPT-6 Astra — Deep Principle’s own figure, not independently verified
Hardware needed Tested with SGLang on 8 GPUs; roughly a 452 GB download for the full BF16 weights
Price No separate price announced; Deep Principle’s cost charts use GLM-5.3 pricing for Hydrogen-1.0

The single most important thing to understand about Hydrogen-1.0 is that it is a specialist, not a generalist. The last two years of AI progress have been dominated by general-purpose models that try to do everything — write, code, reason, chat — reasonably well. the model goes the other way. It takes a strong general model and spends its entire additional training budget on one narrow slice of human knowledge: how scientists actually do science. That bet is what makes the release interesting, and it is also what limits who can use it.

What Is Hydrogen-1.0?

Hydrogen-1.0 is a large language model designed to help with scientific research in three fields: chemistry, materials science and life sciences. Deep Principle describes it as a model that “helps scientists formulate questions, design informative experiments and use the resulting evidence to guide further discovery.”

In plain language, it is meant to sit next to a researcher the way a very knowledgeable lab colleague would — suggesting what to test next, checking whether a reaction plan makes sense, turning a materials question into a simulation workflow, and reading through long documents without losing the thread.

That is a different job from the one most famous AI models do. A general chatbot is judged on how well it answers a question in one turn. A science model is judged on whether its suggestions survive contact with reality: whether the experiment it proposes is actually runnable, whether the reaction it predicts actually happens, and whether the material it recommends actually has the properties it claims.

Deep Principle’s whole pitch is that Hydrogen-1.0 was trained on the kind of evidence that teaches exactly that judgement — not just textbooks and papers, but the records of real experiments run in real laboratories, including the failed attempts that never make it into a published paper.

Hydrogen-1.0 Is a Science Model, Not a Chatbot

It is worth being blunt about what Hydrogen-1.0 is not. It is not a consumer chatbot with a friendly web interface, and Deep Principle does not present it as one. There is no free chat website where you can type a question and get an answer in seconds. The main way to use the model is to download its weights and run them on your own servers, call it through the NeuroSpark API, or use it inside Mira, Deep Principle’s own “AI Scientist” product. That already tells you who the intended audience is: research groups, companies with compute clusters, and developers building science tools — not casual users.

The model is also not a replacement for a scientist. Its own model card warns plainly that its outputs “should not be treated as verified experimental findings” and that it may produce incorrect facts, calculations, equations, references or explanations. That warning matters more here than it does for a general chatbot, because a wrong answer in chemistry can mean a wasted week of lab work, a ruined batch of material, or — in the worst case — a safety incident.

Hydrogen-1.0 is a tool for generating hypotheses and plans that a trained human then checks, not an oracle that hands down answers.

Scientist examining a test tube in a laboratory, the real-world work Hydrogen-1.0 is designed to assist

Who Is Deep Principle?

Deep Principle is a young company from Hangzhou that describes itself as building “intelligence for scientific discovery in chemistry, materials science and life sciences.” Unlike the giant labs that dominate AI headlines, it runs its own research laboratories — and that detail is central to the Hydrogen-1.0 story. The company says it built the model by post-training GLM-5.3 on domain data “together with records and agent trajectories from our own laboratories.” In other words, the training data includes what happened when its own scientists and AI agents actually ran experiments: what they tried, what worked, what failed, and what they decided next.

The company is led by founder and chief technology officer Chenru Duan, who announced the release on X on 2 November 2026 and posted the benchmark comparison against GPT-6 Astra. Deep Principle worked with two partners on the release: UniPat and Mind Lab helped with data, the training pipeline and infrastructure, while NeuroSpark contributed training infrastructure and now hosts the API.

A formal press release followed on 4 November 2026, distributed on 6 November, positioning Hydrogen-1.0 as a tool that helps chemistry and materials research teams “plan experiments, run computational workflows and explore candidates for desired material properties.”

For a small startup, releasing a 753-billion-parameter model under the MIT licence is a bold move. It puts Deep Principle in the same conversation as the biggest open-weight releases of the year — and it is a bet that openness, not secrecy, is the way to win trust with scientists. The question is whether the benchmarks hold up once independent researchers get their hands on the weights. So far, no independent lab has replicated Deep Principle’s numbers, and every figure in this article comes from the company’s own evaluations. Keep that in mind throughout.

How Hydrogen-1.0 Was Built

Hydrogen-1.0 was not trained from scratch. Deep Principle started with GLM-5.3, the open-weight general-purpose model from Z.ai, and then ran an additional training phase — called post-training — to specialise it. That distinction matters, because it explains both the model’s strengths and its limits. Everything the model knows about language, reasoning and general knowledge comes from GLM-5.3. Everything it knows about doing science comes from the extra training Deep Principle added on top.

The post-training data had three ingredients, according to the company’s announcement. First, domain data from chemistry, materials science and life sciences — the textbooks, papers and structured knowledge of those fields. Second, records from Deep Principle’s own laboratories — the day-to-day notes of experiments actually run. Third, “agent trajectories” — records of what AI agents did while working on scientific tasks in those labs, including the sequence of decisions they made and the results they got. The idea is that a model trained on how discovery actually happens, including the dead ends, will be better at guiding discovery than a model trained only on polished, published results.

Training: Supervised Fine-Tuning Plus Reinforcement Learning

Deep Principle says its training pipeline combines two standard techniques. The first is supervised fine-tuning, where the model is shown many examples of good scientific work — questions paired with good answers, experiment plans paired with good outcomes — and learns to imitate them. The second is reinforcement learning, where the model tries things, gets scored on the results, and gradually learns which approaches work and which do not. Used together, the two stages first teach the model what good science looks like, and then push it to actually do science better rather than just describe it.

What Deep Principle did not release is just as important as what it did. The company states plainly in the model card that “full datasets and recipes are not released with the weights.”

That means other researchers can download and use Hydrogen-1.0, but they cannot see exactly what it was trained on, or reproduce the post-training process themselves. This is the same tension that runs through most open-weight releases in 2026: the weights are open, but the data and the recipe stay closed. For a science model, that gap matters more than usual, because knowing what a model was trained on is part of knowing when to trust it.

The training ran on a serious stack: Megatron and SGLang for training and serving, on NVIDIA CUDA and AMD ROCm hardware. Deep Principle says the model was tested running BF16 weights with SGLang across eight GPUs, using tensor parallelism — the work split across all eight cards at once. That is a data-centre setup, not a workstation setup, and it sets the tone for everything about running the model yourself, which we cover in the hardware section below.

Hydrogen-1.0 Architecture and Specs

Under the hood, Hydrogen-1.0 is a mixture-of-experts model, or MoE. That means the model is made up of many smaller “expert” networks, and only a few of them switch on for any given token of text. The Hugging Face repository lists 753 billion total parameters, with about 40 billion active per token, and eight experts per token. The architecture is tagged as glm_moe_dsa, and the base model is listed as zai-org/GLM-5.3. The design is the reason a model this large can be served at all: at any moment, most of its parameters are sitting idle, and only the relevant experts do the work.

Here are the headline specifications of Hydrogen-1.0, as published on its Hugging Face model card:

  • Total parameters: 753 billion (mixture-of-experts, 8 experts per token)
  • Active parameters per token: about 40 billion
  • Base model: Z.ai GLM-5.3
  • Context window: 1 million tokens
  • Maximum output: up to 128,000 tokens
  • Main weights format: BF16 (brain floating point 16-bit)
  • Quantized versions: NVFP4 for NVIDIA Blackwell B200 and B300 GPUs; MXFP4 for AMD Instinct MI350X and MI355X
  • Licence: MIT
  • Serving: SGLang, tested on 8 GPUs with tensor parallelism
  • Download size: around 452 GB for the full BF16 weights (about 10 hours at 100 Mbit/s)

Size: 753B Parameters, 40B Active

The 753-billion figure sounds enormous, and it is — but the number that matters for day-to-day use is the 40 billion active parameters. Because only eight experts fire for each token, the compute cost of generating text is closer to that of a 40-billion-parameter model than a 753-billion one.

The trade-off is memory: all 753 billion parameters still have to live somewhere, which is why the download is roughly 452 GB and the model needs multiple high-end GPUs just to hold the weights. In effect, Hydrogen-1.0 trades cheap thinking for expensive housing. That is a familiar bargain in the MoE world — the same logic sits behind other big open models of 2026 — but it is worth stating plainly, because it decides whether you can run the model at all.

The quantized versions are Deep Principle’s answer to that problem. NVFP4 and MXFP4 compress the weights into 4-bit floating point, shrinking the memory footprint dramatically for the newest NVIDIA Blackwell and AMD Instinct hardware. The company publishes these in separate repositories, so a team with B200, B300, MI350X or MI355X cards can run a much lighter copy of Hydrogen-1.0. Even so, this is still server hardware. There is no realistic path to running the full model on a home GPU, a Mac, or a cloud notebook — and even a heavily quantized copy needs several hundred gigabytes of memory.

Context: 1 Million Tokens In, 128,000 Out

The context window of Hydrogen-1.0 is one million tokens — enough to feed in an entire research literature review, a stack of related papers, or a very long lab notebook in a single prompt, and still have room for the model to reason about them. The maximum output of 128,000 tokens means it can also produce very long answers: a full experiment plan, a multi-step simulation workflow, or a detailed materials report, without being cut off mid-thought.

There is a small-print warning here, though, and it comes from the model card itself: the usable context and output budgets depend on the inference framework, the server configuration and the runtime resources. In other words, one million tokens is what the model was built for, not necessarily what you will get on your hardware. Long-context performance also degrades on real workloads — a model that can technically accept a million tokens does not always use the middle of that window well. For science work, where the crucial detail might sit in the middle of a 400-page document, that is something to test yourself before trusting it.

3D molecular structure illustration showing the chemistry research Hydrogen-1.0 is trained for

Hydrogen-1.0 Benchmarks: Where It Wins and Loses

Every number in this section comes from Deep Principle’s own evaluations, and none of it has yet been independently replicated. That is not a criticism — it is the normal state of affairs a week after any model launch — but it is the frame you should hold in mind while reading. The company has been unusually open about the limits of its own testing, noting where it measured a rival itself and where its internal tests omitted competitors entirely. We have kept those caveats in the text below rather than smoothing them out.

Hydrogen-1.0 vs GPT-6 Astra on FrontierScience-Research

The headline result is on FrontierScience-Research, a benchmark of unfamiliar, multi-step problems in physics, chemistry and biology — the kind of test that asks a model to do research, not just recall facts. Hydrogen-1.0 scored 38.3%, against 33.3% for OpenAI’s GPT-6 Astra at its maximum thinking setting. That is a five-point edge on a hard test where even the best models solve barely more than a third of the problems.

Two details sharpen the picture. First, the comparison is against GPT-6 Astra at maximum thinking, which is the strongest setting of the strongest general model Deep Principle chose to compare against — so the win is not against a weakened rival.

Second, the cost picture is striking: Deep Principle estimates that answering one FrontierScience-Research problem costs about $0.25 in output tokens for Hydrogen-1.0, against $0.88 for GPT-6 Astra — roughly 72% less. That cost figure uses GLM-5.3 pricing for the model, since Deep Principle has not announced a separate price for the model, and it counts output tokens only. Even with that caveat, a cheaper model that also scores higher is exactly the combination that gets research labs’ attention.

It is also worth noting what 38.3% means in absolute terms. On this benchmark, the model still fails nearly two out of three problems. The story is not that Hydrogen-1.0 has solved scientific research; it is that a specialist model, built for a fraction of the cost of a frontier general model, now outperforms that general model on the specialist’s home turf. That is a meaningful result, but it is a relative one.

GPQA Diamond: Roughly Matching Kimi K3

The second headline number is on GPQA Diamond, a graduate-level science benchmark built from difficult questions written by domain experts. Hydrogen-1.0 scored 93.0%, which Deep Principle says roughly matches Kimi K3, one of the strongest open models of 2026, and improves over its own starting point, GLM-5.3. For context, Kimi K3 is the model behind Moonshot AI’s open-weight push, and matching it on a graduate science test is a serious claim for a first release from a small startup.

As with the FrontierScience result, the number is self-reported and has not been independently re-run on the released weights. GPQA Diamond is also a multiple-choice benchmark, which tests knowledge and reasoning under exam conditions rather than the messy, open-ended work of a real lab. A high GPQA score tells you the model knows a lot of science. It does not, by itself, tell you the model can plan a good experiment — which is why Deep Principle’s internal lab-task tests, covered next, are arguably more interesting than either headline number.

Lab Tasks: Computing, Reactions and NMR

Beyond the public benchmarks, Deep Principle ran three internal scenario tests designed to look like real lab work. On two of them, Hydrogen-1.0 beat GPT-6 Astra; on the third, it lost. The company reports that at maximum thinking, the model beats GPT-6 Astra by 5.13 percentage points on scientific computing and by 4.76 points on forward reaction prediction — predicting what a chemical reaction will produce. On the third task, NMR structure elucidation — working out a molecule’s structure from its nuclear magnetic resonance spectrum — the model was weaker than GPT-6 Astra.

That mixed result is actually reassuring. A company that reports its losses as well as its wins is easier to trust than one that only publishes victories, and the pattern makes sense: forward reaction prediction and scientific computing are exactly the tasks that benefit from training on lab records and simulation workflows, while NMR elucidation is a specialist skill that may need different data. If you work in NMR, this is a clear signal to test the model carefully before relying on it. If you work in reaction planning or computational chemistry, the numbers suggest Hydrogen-1.0 is worth a serious look.

There is one more internal test worth mentioning, because it shows how Deep Principle thinks about evaluation. In a 117-question experiment-planning comparison, the company omitted GPT-6 Astra and Anthropic’s Opus 5.5 entirely — because, by its own account, those two models rejected more than half of the requests. That is a striking detail: the frontier models refused to answer a large share of experiment-planning questions, presumably on safety grounds, while the specialist model answered them.

It raises a genuine question for the field: when a general model refuses a legitimate scientific question, is that safety working as intended, or is it a capability gap that specialist models will fill? Deep Principle clearly believes it is the latter.

Materials Discovery: The 4.318 eV Result

The most eye-catching test in the announcement is the materials-discovery experiment. Deep Principle asked the models to search for materials with desired properties over five rounds, measuring the “best-so-far gap” — essentially, how close the best material found so far is to the target. By round five, the mean best-so-far gap reached 4.318 eV for Hydrogen-1.0, compared with 3.689 eV for GPT-6 Astra and 3.353 eV for Kimi K3. In other words, after five rounds of searching, the specialist model had found better materials than either frontier rival.

This is the result that best captures what Hydrogen-1.0 is for. Materials discovery is a search problem: you start with a target property, propose candidates, test them (in simulation, at first), learn from the results, and propose better candidates. A model trained on lab records and discovery trajectories — on the process of searching, not just the facts — should be good at exactly this. The 4.318 eV figure is Deep Principle’s own, measured on its own test, and it needs independent confirmation. But if it holds up, it is the strongest argument yet that specialist post-training is a real shortcut to capability in science, not just a marketing story.

The Cost Angle: Why the Model Is Cheaper to Run

Cost is the quiet second headline of the Hydrogen-1.0 release. Deep Principle’s charts put the model at about $0.25 of output-token cost per FrontierScience-Research problem, against $0.88 for GPT-6 Astra — a saving of roughly 72%. The reason is structural: the model is priced at GLM-5.3 rates in the company’s comparisons, and GLM-5.3 is an open-weight model whose serving costs are far below those of a closed frontier API. For a research lab running thousands of queries — screening candidate materials, planning dozens of experiments, reading hundreds of papers — that gap compounds fast.

Be careful with the exact numbers, though. The $0.25 figure is an estimate based on output tokens, and the pricing is borrowed from the base model because Deep Principle has not published a Hydrogen-1.0 price of its own. If you run the model yourself, your real cost is hardware, electricity and engineering time, not tokens. If you use the NeuroSpark API, the price will be whatever NeuroSpark charges. The honest way to read the cost claim is directional: a specialist open-weight model should be substantially cheaper per scientific answer than a closed frontier model — and on Deep Principle’s own measurements, it is.

Why the Numbers Need a Caveat

To be fair to Deep Principle, the company has not hidden the limits of its evidence. Its own materials state that all reported figures come from its evaluations, that the FrontierScience figure for GPT-6 Astra was measured by the company itself, and that no independent results had been published at the time of writing. The benchmarks are also run on the company’s own test sets in some cases, and the cost accounting uses the base model’s pricing. None of that makes the numbers wrong — but it does mean they are claims awaiting verification, not established facts.

What would settle the question is independent testing on the released weights: another lab downloading Hydrogen-1.0, running FrontierScience-Research and GPQA Diamond under standard conditions, and publishing the results. That is precisely what open weights make possible, and it is the real test of the release. Until then, the sensible position is the one Deep Principle itself invites: the model looks strong on its own tests, the design is sound, and the proof will come from the community.

How to Access Hydrogen-1.0

Deep Principle offers four routes to Hydrogen-1.0, and they serve very different users. Picking the right one depends on whether you want to run the model, call it, or simply use it inside a finished product.

Download: Hugging Face and ModelScope

The most direct route is to download the weights yourself from Hugging Face, at deep-principle/Hydrogen-1.0, or from ModelScope. The weights are published under the MIT licence, which permits use, modification and commercial use, subject to the licence’s terms — including keeping the licence notice attached. That is about as permissive as software licensing gets, and it is a deliberate choice: Deep Principle wants research groups and companies to build on the model without legal friction.

There is one condition on the download, and it surprises some people. The Hugging Face repository is gated: before you can download, you are asked to agree to share your Hugging Face account information — your username and email address — with Deep Principle. That is not a payment, and it is not a restrictive licence, but it is a form of access control, and it means the download is not fully anonymous.

If that bothers you, the ModelScope mirror is the alternative route to the same weights. Either way, budget for the size: the full BF16 download is around 452 GB, which takes about ten hours on a 100 Mbit/s connection — and much longer on a slower one.

API Access: NeuroSpark

If you do not want to run the model yourself, the NeuroSpark API is the hosted route. NeuroSpark, which also contributed training infrastructure to the release, offers Hydrogen-1.0 through an API under the model ID hydrogen-1.0, with an OpenAI-compatible interface and a reasoning_effort setting that lets you control how hard the model thinks on each request. That compatibility matters: if your tools already talk to OpenAI-style APIs, switching the endpoint and model name is often all it takes to try the model.

Deep Principle has not published a separate API price for Hydrogen-1.0, and its own cost charts use GLM-5.3 pricing as the stand-in. That means the real API price is whatever NeuroSpark sets, and you should check it before building anything serious on top of it. For a quick evaluation — a few dozen experiment plans, a materials screen, a literature summary — the API is by far the cheapest way to find out whether the model is useful for your work, because you pay nothing for hardware.

Mira: Deep Principle’s AI Scientist

The fourth route is Mira, Deep Principle’s own AI Scientist product, where Hydrogen-1.0 is the engine inside a finished application. Mira is aimed at chemistry and materials research teams who want the model’s capabilities — experiment planning, computational workflows, candidate search — without managing infrastructure or writing API code. Deep Principle’s press release positions Mira as the way “researchers can use the model” alongside the downloads and the API, and it is the route most likely to suit a working lab that just wants results.

There is not much public detail yet about Mira’s pricing, availability or exact feature set, and the launch materials do not include a detailed roadmap for it. That is normal for a first release: the weights and the API are the concrete deliverables, and Mira is the product built on top of them. If you are evaluating Mira, ask for a trial against your own lab’s problems — a demo on generic examples tells you far less than a test on the experiments you actually run.

Rows of GPU server racks in a data centre, the hardware Hydrogen-1.0 needs to run

Hardware: Can You Run Hydrogen-1.0?

The honest answer for most readers is no — not the full model, not on hardware you own. Hydrogen-1.0 is a data-centre model, and Deep Principle does not pretend otherwise. The company tested it with SGLang on eight GPUs, using tensor parallelism to split the work across all eight cards. That is the reference deployment, and it is the setup you should assume if you plan to self-host.

Hardware: What 8 GPUs Really Means

Eight GPUs, in this context, means eight high-end data-centre accelerators — the class of card that costs thousands of dollars each and lives in a rack, not on a desk. The full BF16 weights are around 452 GB, so the cards need enough combined memory to hold them, plus room for the context window and the model’s working memory during generation. Even the memory math is unforgiving: a single consumer GPU with 24 GB of memory holds roughly a twentieth of what the model needs.

The serving stack matters too. Deep Principle’s tests used SGLang, a fast serving framework for large models, with support for NVIDIA CUDA and AMD ROCm. Running Hydrogen-1.0 is not just a matter of downloading the weights and typing a command; it is a small infrastructure project, with GPU drivers, serving software, memory management and monitoring to set up and keep running. A research group with a cluster and an engineer to run it can do this. A lone researcher with a laptop cannot — and should use the API or Mira instead.

Quantized Versions: NVFP4 and MXFP4

The quantized versions are the one concession to smaller setups. Deep Principle publishes NVFP4 weights for NVIDIA’s newest Blackwell cards (the B200 and B300) and MXFP4 weights for AMD’s Instinct MI350X and MI355X. Four-bit quantization compresses each weight to a quarter of its BF16 size, so the model fits in far less memory — but the hardware is still the newest, most expensive data-centre silicon on the market. These versions make Hydrogen-1.0 cheaper to serve for the teams that already own that hardware; they do not make it runnable on older or consumer cards.

The practical takeaway is simple. If you have a cluster, the weights are yours to run under the MIT licence. If you do not, the API is the sensible route, and it is the route Deep Principle clearly expects most users to take. There is no shame in that: the same is true of almost every large open-weight model of 2026, and the API exists precisely so that the model’s capabilities are not limited to the few dozen labs with eight-GPU servers.

Hydrogen-1.0 Licence and Limits

The licence is one of the most attractive parts of the Hydrogen-1.0 release — and one of the most misunderstood. Here is what it actually allows, and what it does not.

MIT Licence: What It Allows

The weights of Hydrogen-1.0 are released under the MIT licence, one of the most permissive licences in software. In practice, that means you can use the model for research or commercial purposes, modify it, fine-tune it, and build products on top of it, as long as you include the licence notice. There is no royalty, no field-of-use restriction in the licence itself, and no requirement to publish your own changes. For a company that wants to build a science tool on top of a frontier-class model without paying API fees or negotiating a commercial licence, that is a genuinely valuable offer.

The MIT licence applies to the weights — the trained parameters of the model. It does not apply to the base model’s original training data, or to the lab records Deep Principle used in post-training, which remain the company’s property. That is a standard distinction, but it is worth stating, because “open-weight” is sometimes misread as “everything is open.” It is not. The model is open; the data behind it is not.

What Is NOT Open

The closed parts of the release are the full training datasets and the training recipes — the exact mix of data, the hyperparameters, and the step-by-step process that turned GLM-5.3 into Hydrogen-1.0. Deep Principle says so directly in the model card: “full datasets and recipes are not released with the weights.”

That has two consequences. First, nobody else can reproduce the training exactly, which limits how much the community can verify or extend the work. Second, nobody can audit the training data for gaps or biases — for example, whether the lab records cover some kinds of chemistry much better than others. For a science model, where the training data is the product as much as the weights are, that is a real limitation, and it is the main reason independent verification matters so much.

The Model Card Warning

The model card carries a warning that every user should read before trusting an output. It says the model may produce incorrect facts, calculations, equations, references or explanations, and that its outputs “should not be treated as verified experimental findings or as grounds for decisions affecting health, safety or other consequential outcomes.” That is not boilerplate.

A language model that is wrong about a historical fact is an embarrassment; a science model that is wrong about a reaction, a dose, or a material property can be dangerous. Treat Hydrogen-1.0 as a hypothesis generator whose outputs need checking by a qualified human, with real experimental validation, before anything is acted on. That is how Deep Principle itself frames it, and it is the only safe way to use it.

Who Should Use Hydrogen-1.0?

Hydrogen-1.0 is a specialist tool, and like any specialist tool, it is excellent for some people and a waste of time for others. Here is a straight answer for each group.

A Good Fit For:

  • Chemistry and materials research groups with access to GPU clusters, who want an open-weight model they can fine-tune on their own data and run without per-token API bills.
  • Computational chemistry teams doing reaction prediction, simulation workflows and materials screening, where the model’s reported strengths lie.
  • Pharma and biotech researchers working in life sciences, who need long-context reading of papers and experiment planning — with human verification of every output.
  • AI developers building science tools, who can call the NeuroSpark API or build on the MIT-licensed weights without negotiating a commercial licence.
  • University labs and open-science projects that need a strong science model they can study, modify and share under a permissive licence.

NOT a Good Fit For:

  • Casual users looking for a chatbot. There is no free chat website, and the model is not designed for general conversation.
  • Anyone without GPU infrastructure who insists on self-hosting. Use the API or Mira instead — the full model needs eight data-centre GPUs.
  • NMR specialists, for now: the model’s own tests show it is weaker than GPT-6 Astra on NMR structure elucidation, so test it carefully in that area.
  • Anyone who needs verified answers without human checking. The model card warns explicitly against treating outputs as verified findings, especially where health or safety is involved.
  • Teams that need full reproducibility of the training process. The datasets and recipes are not released, so you cannot rebuild the model from scratch.

Research team discussing data on a computer in a laboratory, the audience Hydrogen-1.0 is built for

Hydrogen-1.0 vs GPT-6 Astra vs Kimi K3

Deep Principle’s announcement is built around comparisons, so here they are in one place — with the source of each number, so you can judge them fairly.

Hydrogen-1.0 GPT-6 Astra Kimi K3
Developer Deep Principle OpenAI Moonshot AI
Type Open-weight specialist (science) Closed frontier generalist Open-weight generalist
Parameters 753B total, ~40B active Not disclosed Not disclosed in this comparison
Licence MIT (weights) Proprietary API Open-weight
FrontierScience-Research 38.3% 33.3% Not reported here
GPQA Diamond 93.0% Not reported here Roughly matched by Hydrogen-1.0
Cost per FSR problem* ~$0.25 ~$0.88 Not reported here
Materials discovery (round 5) 4.318 eV 3.689 eV 3.353 eV
Context window 1M tokens Not disclosed here Not disclosed here

*Cost figures are Deep Principle’s estimates using GLM-5.3 pricing for the model and output tokens only. All benchmark figures are vendor-reported and not yet independently verified.

The table tells a clear story. On science-specific tests, the specialist beats the generalist — and does it more cheaply. That is the entire thesis of Hydrogen-1.0 in one sentence. But the table also shows the limits of the comparison: Kimi K3’s FrontierScience score is not reported, GPT-6 Astra’s GPQA score is not reported, and the cost figures are estimates. A full, independent comparison on identical conditions is still missing, and it is the single most useful thing the community could produce in the coming weeks.

For readers of this site, two of the rivals are familiar. Kimi K3 is the open-weight model we covered in our Kimi K3 guide, and the open-weight ecosystem it belongs to is the same one our ERNIE 5.5 Flash Base guide and our DeepSeek V4.1 Flash guide describe. Hydrogen-1.0 is best understood as the science-specialist branch of that same open-weight family tree — the same roots, a different job.

What Hydrogen-1.0 Means for Science AI

Hydrogen-1.0 arrives at a moment when the AI industry is asking a hard question: is the future bigger general models, or smaller specialist ones? For two years, the answer looked like “bigger.” Then the open-weight wave of 2026 changed the economics.

Models like Kimi K3, ERNIE 5.5 Flash Base and DeepSeek’s V4.1 family showed that openly available models could match or beat closed ones on specific tasks, at a fraction of the cost. the model pushes that logic one step further: take a strong open model, specialise it hard on one domain, and beat the frontier generalist on that domain’s own tests.

If the results hold up under independent testing, the implications are significant. First, it would validate post-training as a strategy: you do not need to train a frontier model from scratch to beat one — you need a good base model and the right data.

Second, it would shift value toward whoever owns the best domain data. Deep Principle’s edge is not compute or scale; it is the lab notebooks, the experiment records, and the discovery trajectories that only a company running its own labs can collect. Data like that cannot be scraped from the internet, and it cannot be bought off the shelf. Third, it would give research labs a real alternative to closed APIs: a model they can download, fine-tune, inspect and run on their own terms, under a licence that lets them build a business on it.

There are reasons for caution, too. The benchmarks are self-reported, the datasets are closed, and the model is too big for most labs to run. The NMR weakness is a reminder that specialisation is uneven — a model can be excellent at reaction prediction and mediocre at structure elucidation, and a researcher who does not know that will be misled. And the experiment-planning refusal gap — where frontier models declined to answer questions the specialist model answered — is a genuine safety debate, not a settled victory. Some of those refusals may be exactly right, and a model that answers everything is not automatically a better model.

The fairest summary is this: Hydrogen-1.0 is the most interesting science-model release of the autumn, with a credible design, a genuinely novel training story, and benchmark claims that deserve independent testing rather than instant belief or instant dismissal. Watch for three things in the coming weeks: independent benchmark runs on the released weights, reports from labs that try it on their own problems, and whether Deep Principle releases any of the training data or a smaller, more runnable version. Any one of those would tell us more than the launch announcement did.

Frequently Asked Questions About Hydrogen-1.0

What is Hydrogen-1.0?

Hydrogen-1.0 is a 753-billion-parameter open-weight AI model built by Deep Principle for scientific research in chemistry, materials science and life sciences. It was created by post-training Z.ai’s GLM-5.3 on scientific domain data plus records and agent trajectories from Deep Principle’s own laboratories, and it was released on 2 November 2026 under the MIT licence.

Is it free to use?

The weights are free to download under the MIT licence, which permits commercial use, modification and redistribution with the licence notice attached. But “free” does not mean “cheap to run”: the full model needs eight data-centre GPUs, and the download is around 452 GB. The NeuroSpark API is the hosted alternative, and its price is set by NeuroSpark — Deep Principle has not announced a separate price for Hydrogen-1.0.

Can I run it on my own computer?

Realistically, no. The full BF16 model needs eight high-end GPUs and roughly 452 GB of storage for the weights alone, and Deep Principle tested it with SGLang across eight cards. Even the quantized NVFP4 and MXFP4 versions need the newest NVIDIA Blackwell or AMD Instinct data-centre hardware. A home GPU, a Mac or a cloud notebook cannot run it. Use the NeuroSpark API or Deep Principle’s Mira product instead.

Does it really beat GPT-6 Astra?

On Deep Principle’s own tests, yes — Hydrogen-1.0 scored 38.3% on FrontierScience-Research against 33.3% for GPT-6 Astra, and beat it on scientific computing and forward reaction prediction. But those are the company’s own figures, measured by the company itself, and no independent lab has replicated them yet. The model was also weaker than GPT-6 Astra on NMR structure elucidation. Treat the headline as a claim awaiting verification, not an established fact.

What is the context window?

Hydrogen-1.0 supports a context window of one million tokens, with up to 128,000 output tokens. The model card notes that the usable budgets depend on your inference framework, server configuration and runtime resources, so the full million tokens is a design figure rather than a guarantee on every setup.

What licence is it released under?

The weights are released under the MIT licence, which is highly permissive: you can use, modify and commercialise the model, provided you keep the licence notice. The download from Hugging Face is gated — you agree to share your Hugging Face account information before downloading. The full training datasets and recipes are not released.

How do I access it?

There are four routes: download the weights from Hugging Face (deep-principle/Hydrogen-1.0) or ModelScope, call the model through the NeuroSpark API (model ID hydrogen-1.0, OpenAI-compatible), or use it inside Mira, Deep Principle’s AI Scientist product. For most people, the API is the fastest way to try it.

Is it safe to use for real lab work?

Only with human verification. The model card warns that Hydrogen-1.0 may produce incorrect facts, calculations, equations, references or explanations, and that its outputs should not be treated as verified experimental findings or as grounds for decisions affecting health, safety or other consequential outcomes. Use it to generate hypotheses and plans, then check everything with a qualified scientist and real experiments.

Who made it?

Hydrogen-1.0 was made by Deep Principle, a materials-research AI company based in Hangzhou, China, led by founder and CTO Chenru Duan. The company runs its own laboratories, and the model was trained on data from those labs. UniPat and Mind Lab helped with data, the training pipeline and infrastructure, and NeuroSpark contributed training infrastructure and hosts the API.

What is Mira, Deep Principle’s AI Scientist?

Mira is Deep Principle’s finished product built on Hydrogen-1.0, aimed at chemistry and materials research teams. It offers the model’s capabilities — experiment planning, computational workflows, materials candidate search — without requiring the user to manage GPUs or write API code. Public details on pricing and availability are still limited.

The Bottom Line on Hydrogen-1.0

Hydrogen-1.0 is the most interesting science-model release of the autumn for a simple reason: it is a specialist that beats a generalist on the specialist’s own ground, and it does it openly. A 753-billion-parameter model, built by post-training GLM-5.3 on real lab records, released under the MIT licence, scoring 38.3% against GPT-6 Astra’s 33.3% on FrontierScience-Research at roughly a quarter of the output-token cost — that is a release worth paying attention to, whatever you think of the individual numbers.

But it is also a release with clear limits. The benchmarks are self-reported and unverified. The training data and recipes are closed. The model is too big for almost anyone to run, and its own card warns against treating its outputs as verified findings. The NMR weakness and the experiment-planning refusal gap are reminders that the story is more complicated than the headline.

If you run a research lab with a GPU cluster, download Hydrogen-1.0 and test it on your own problems — that is what open weights are for. If you are a researcher without that infrastructure, try it through the NeuroSpark API or Mira, and judge it on your own experiments, not on a press release.

And if you are simply watching the AI industry, watch what happens next: independent benchmarks, lab reports, and whether Deep Principle opens up any more of the training story. The launch is the beginning of the test, not the end of it. For the official announcement, see Deep Principle’s own post, Introducing Hydrogen-1.0, and the weights on the Hydrogen-1.0 Hugging Face page.

Leave a Comment

Your email address will not be published. Required fields are marked *