Hydrogen-1.0 Review: Deep Principle's Free 753B Science Model — and the $53,280 Catch

A 753-billion-parameter AI model for chemistry and materials science just went open weight, free, under the MIT license. If that sounds too good to be true, you are reading the right article. Deep Principle released Hydrogen-1.0 on November 2, 2026, and the weights really are $0. The bill shows up somewhere else: you need eight datacenter GPUs to run it, which lists at about $25,920 a month on Hugging Face's own price page — $53,280 if you want the newest Blackwell cards the model was quantized for.

This is the honest walkthrough: what Hydrogen-1.0 is, what it can actually do in a lab, what its benchmarks prove and what they don't, and exactly what "free" costs in dollars, pounds, and Canadian dollars.

What Hydrogen-1.0 actually is

Hydrogen-1.0 is a scientific research model from Deep Principle, a Hangzhou AI-for-materials company founded in 2024 by MIT PhDs Haojun Jia and Chenru Duan that has raised over $100 million and was named to the World Economic Forum's 2025 Technology Pioneers cohort, according to the company's October 29 announcement. It targets chemistry, materials science, and life sciences: formulating research questions, planning experiments, analyzing results, and tackling open-ended discovery tasks.

It is not a model trained from scratch. Deep Principle took Z.ai's open GLM-5.3 and post-trained it with supervised fine-tuning and reinforcement learning on three things most labs never share: curated scientific data, their own electronic laboratory notebook records (what the researcher wanted to learn, the protocol, the observations, the outcome, including failures), and full discovery trajectories generated with the company's Mira agent. That last ingredient is the interesting one. A lab notebook tells a model how an experiment ended. A discovery trajectory tells it how a scientist decided what to try next after the last three things failed. The company argues that distinction is what turns a chatbot into a research assistant, and it is the core bet behind the release.

The architecture is a mixture-of-experts design with 753 billion total parameters and roughly 40 billion active per token, inherited from the GLM-5.3 base. It supports a 1-million-token context window and up to 128,000 output tokens, so it can ingest full research papers, long protocols, or genomic-scale sequences in one pass. Three weight packages ship together: unquantized BF16, NVFP4 for NVIDIA's Blackwell B200/B300 cards, and MXFP4 for AMD's Instinct MI350X/MI355X, per the model-card details reported at launch.

What it does in practice

The launch post walks through four scenarios the company's own research team runs into, and they read less like chatbot demos and more like lab work.

First, simulation workflows. Hydrogen-1.0 takes a materials question about a polyethylene melt, writes Python to build it, runs GROMACS, detects a topology error, repairs it, and restarts the simulation, then explores the corrected 40-nanosecond trajectory for density, molecular packing, diffusion, and chain dimensions. Pre-trained tool use matters here: the model ships knowing how to write executable scripts, handle GROMACS topology checks, and emit JSON experiment protocols, MarkTechPost reported.

Second, reaction networks. Given an Ireland–Claisen rearrangement, the model explores a network of 27 molecular structures across 30 plausible reactions, including competing branches and return paths through Cope and Claisen rearrangements and silyl shifts, and compares electronic barriers step by step to flag bottlenecks.

Third, spectra. It starts with seven possible monoformylation products of 2-bromonaphthalene and works through a supplied proton NMR trace, inspecting downfield signals and the crowded aromatic region before committing to a structure.

Fourth, and most unusual, lab compilation. Hydrogen-1.0 writes a JSON protocol for a 12-vial Suzuki coupling study, then rehearses it in a virtual environment that checks station capabilities, dependencies, and action order. The rehearsal catches an attempt to aspirate without a fresh pipette tip; the model fixes the action order and compiles the corrected protocol before anything touches real reagents. Catching the error in simulation instead of at the bench is the company's answer to safety questions about giving agents more autonomy in labs.

The benchmarks, with the fine print

Here is where the article earns its honesty badge. Every number below comes from Deep Principle's own evaluations, published in the launch post. None has been independently verified.

Two catches deserve emphasis. First, the headline 38.3% vs 33.3% on FrontierScience-Research does not appear in the company's blog post; it comes from launch coverage quoting co-founder Chenru Duan, and Tokenstrom notes the GPT-6-Astra figure was measured by Deep Principle itself, not by OpenAI or a third party. Second, the post's own cost-versus-performance charts price Hydrogen-1.0's API cost using GLM-5.3's rates as a stand-in, because Hydrogen-1.0 has no published per-token price of its own. That is disclosed in the figure captions, and it tells you something: the company has not yet said what a hosted query will cost.

The real price of "free"

The weights are genuinely free. MIT license, no royalties, no seat fees, commercial use allowed. The Hugging Face repo deep-principle/Hydrogen-1.0 is gated, which in practice means a free Hugging Face account and one click to accept the license before downloading. As of November 4 the repo showed about 11,800 downloads and 220 likes, respectable for a two-day-old specialist model.

Then the physics of 753 billion parameters arrives. The BF16 checkpoint is a 452 GB download, roughly ten hours on a 100 Mbit/s connection, per launch coverage. The reference deployment runs SGLang across eight GPUs with tensor parallelism. You cannot run this on a laptop, a Mac, or a free Colab tier; even heavily quantized, it needs several hundred gigabytes of memory.

So what does eight datacenter GPUs actually cost? Hugging Face's own Inference Endpoints price list, current as of November 2026, gives the cleanest public numbers:

Currency conversions use November 4, 2026 mid-market rates (1 USD = £0.74 = C$1.40) and are approximate. The B200 row is the honest reference point: Deep Principle ships a dedicated NVFP4 checkpoint for Blackwell cards and tested the unquantized model on eight GPUs, so budgeting for anything less is wishful thinking. An always-on endpoint is the expensive way to run it; batch jobs and spot capacity would cut the bill, but the company publishes no guidance on either.

The alternative to self-hosting is the hosted API through NeuroSpark, Deep Principle's training-infrastructure partner. The announcement lists API access, NeuroSpark's site is live, but there is no pricing page and no published per-token rate for Hydrogen-1.0 anywhere. Until that number exists, the honest price of a hosted query is "unknown," not "cheap." The fourth route, Mira, the company's AI Scientist platform, is the managed option for labs that want the agent layer rather than raw weights; pricing is likewise undisclosed.

How to get it: the four routes

Five things to know before you download

1. The training data and recipes are closed. MIT covers the weights, not the lab notebooks or the RL pipeline that made them valuable. Reproducibility is partial by design: you can inspect and fine-tune the model, but you cannot rebuild it.

2. Every benchmark is company-run. There is no independent evaluation of Hydrogen-1.0 yet. Treat the Astra comparisons as a vendor claim until a third party reruns them. For context on how seriously to take model-line comparisons, see our earlier piece on OpenAI's reasoning-model lineup, which covers GPT-6-Astra's own launch claims.

3. It loses on NMR. Astra at maximum thinking still beats Hydrogen-1.0 on NMR structure elucidation by 3.33 points in the company's own tests. If your work is spectra-first, the "open" model is not the better one on the vendor's own scoreboard.

4. It is a derivative, not a foundation model. The 753B parameters and the architecture come from GLM-5.3; Deep Principle's contribution is the scientific post-training. That is a legitimate and increasingly common pattern: open Chinese foundation models are becoming the starting point for domain specialists. But it also means the model's general capabilities are bounded by its base.

5. The model card warns you itself. The card cautions that outputs may contain incorrect facts, calculations, equations, references, or scientific explanations, and advises verifying claims through primary literature, independent calculations, or experiments. For a model aimed at working scientists, that warning is the most honest paragraph in the whole release.

Who it is for, and who should skip it

Hydrogen-1.0 is for research groups and R&D teams that already run GPU infrastructure and want an inspectable, fine-tunable science specialist without per-token bills or a vendor looking over their shoulder. Computational chemistry and materials groups are the obvious fit; the GROMACS and reaction-network demos are aimed squarely at them.

Everyone else should wait. If you don't have eight datacenter GPUs or a cloud budget in the five figures, the weights are a 452 GB paperweight. Individual researchers and students get more from the coming hosted API, once NeuroSpark publishes a price, or from generalist models like Kimi K3.1 Pro, which we covered at launch and which Hydrogen-1.0 only roughly matches on GPQA-Diamond. This week's foundation-model launches are a reminder of how fast the open-weights leaderboard moves; a two-day-old release is not a long-term commitment.

Is Hydrogen-1.0 really free?

The weights are free under the MIT license, including for commercial use. Everything around the weights costs money: the 452 GB download bandwidth, the storage, and above all the eight datacenter GPUs needed to serve it, which run $14,400–$53,280 a month on Hugging Face's public price list.

Can I run Hydrogen-1.0 on my laptop or Mac?

No. The reference deployment needs eight datacenter GPUs with tensor parallelism, and even the quantized checkpoints need several hundred gigabytes of memory. A laptop, a Mac, or a free cloud tier cannot load it. The practical route for individuals is the hosted NeuroSpark API or the Mira platform once pricing is public.

How does it compare to GPT-6-Astra?

In Deep Principle's own tests, Hydrogen-1.0 beats Astra on FrontierScience-Research (38.3% vs 33.3%, per coverage citing the developer) and on two of three internal research tasks, but Astra still leads on NMR structure elucidation. None of these figures has been independently verified, and Astra's benchmark score was measured by Deep Principle itself.

What does the MIT license actually allow?

Use, modify, and distribute the weights for any purpose, including commercial products, with only the copyright notice preserved. What it does not include: the training datasets, the RL recipes, or any warranty. The model card explicitly warns that outputs may contain errors.

Where do I download Hydrogen-1.0?

From the official Hugging Face repo deep-principle/Hydrogen-1.0 (free account, accept the license), the ModelScope mirror, or via the NeuroSpark API and Mira AI Scientist for hosted access. All four routes are listed in the official launch announcement.

Hydrogen-1.0 is a noteworthy release: the first serious attempt at an open, inspectable AI scientist trained on real lab experience rather than textbooks alone. But "open weights" is not "free AI." Until there is independent verification of the benchmarks and a published API price, the honest summary is a free 753-billion-parameter download with a five-figure monthly hardware bill attached. If your lab already owns the GPUs, download it today. If not, wait for the price list.

This is general information, not procurement or legal advice.

Sources

• Deep Principle, "Introducing Hydrogen-1.0," official blog, November 2, 2026 — blog.deepprinciple.com/introducing-hydrogen-1-0/
• Hugging Face, deep-principle/Hydrogen-1.0 model repository (MIT license, gated) — huggingface.co/deep-principle/Hydrogen-1.0
• Hugging Face, Inference Endpoints pricing (8× GPU monthly rates), read November 4, 2026 — huggingface.co/pricing
• MarkTechPost, "Deep Principle Releases Hydrogen-1.0," November 2, 2026 — marktechpost.com
• Neuralbrew, "Deep Principle Releases Hydrogen-1.0," November 2, 2026 — neuralbrew.ai
• Oossa, "Deep Principle Releases Hydrogen-1.0 Scientific Model," November 2, 2026 — oossa.com
• AIWeekly, "Deep Principle ships 753B science model Hydrogen-1.0 under MIT," November 4, 2026 — aiweekly.co
• i10x.ai, "Deep Principle Releases Hydrogen-1.0 Scientific Model," November 2, 2026 — i10x.ai
• Tokenstrom, "Hydrogen-1.0: 753B open-weight chemistry and materials model under MIT," November 2, 2026 — tokenstrom.com
• Access Newswire, "Inside Deep Principle," October 29, 2026 (company background) — accessnewswire.com
• Exchange-rates.org mid-market rates, November 4, 2026 (1 USD = £0.74 = C$1.40, approximate)