On November 2, 2026, Deep Principle released its first general-purpose large model, Hydrogen-1.0, and opened the weights on the same day. It is a 753B science model post-trained on GLM-5.3, aimed at chemistry, materials science, and the life sciences. The ModelScope model card lists the license as MIT, with the full BF16 weights coming to 753.33B parameters of Safetensors files, roughly 1.51TB, a 1 million token context window, and a maximum output of 128,000 tokens. You can pull everything from the Hydrogen-1.0 model repository on ModelScope, which also hosts NVFP4 quantized builds for NVIDIA Blackwell and MXFP4 builds for AMD Instinct, with Hugging Face mirroring the release.
Open-weight LLMs are everywhere by now, but an open-weight science model in its own tier is rare. This article takes it apart: where science models actually differ from general models, why the training data includes failed experiments, how to read the official benchmark scorecard, and what you can realistically do with a 1.51TB download.
Why science models deserve their own tier
First, clear up a common misunderstanding: a science model is not "a general model that answers science questions better." Ask any flagship general model a chemistry question and it will produce a plausible answer, but real research workflows are not question-and-answer pairs. According to Deep Principle's official tech blog, Hydrogen-1.0 is positioned to help researchers analyze scientific problems, write simulation programs, debug computational anomalies, analyze reaction pathways, and interpret spectra. Look at those verbs: writing, debugging, analyzing, interpreting. They are all process work, not one-shot answers.
The contrast is easiest to see side by side. A general model in answer mode: you ask "what is the product of this reaction," it outputs a product name, done. A science model in research mode: you give it a materials target, it plans the experiment, refines the plan into executable operation instructions, then uses each round of evaluation results to keep recommending the next batch of material candidates, looping as it goes. The first is search-engine-style knowledge retrieval; the second is closer to a colleague running the project with you. Writing code for molecular simulations, figuring out why a calculation crashed, reading NMR peaks back into a structure, screening ligands for metal complexes round after round -- chained together, these are what a research pipeline looks like.
Zoom out one more level, and AI-driven scientific discovery has already been demonstrated by the closed-source camp. The NetEase column AI Yilei put together a useful comparison: OpenAI used agents to prove a finite-time singularity for the Navier-Stokes equations, and Anthropic found a CRISPR-like repeat-sequence enzyme system. Both are landmarks of AI participating in real discovery. The problem is that both proven paths are locked inside closed systems -- no weights to download, no way to reproduce, no way to build on top. Hydrogen-1.0's significance is repaving a road that has already been shown to work, this time with open weights. What you download is a runnable 753B, not a press release.
That is also the logic behind the statement from Deep Principle founder and CTO Duan Chenru: the choice to open the model weights is meant to help researchers understand AI and keep innovating, pushing an open and shared scientific ecosystem. Science has collaborated for decades through papers plus open toolchains; model weights will likely become the next link in that chain.
Inside the training ledger: lab records and failed experiments
The training data composition is the most distinctive part of this release. The ModelScope model card and official channels describe roughly four ingredients: domain knowledge in chemistry, materials, and the life sciences; feedback from research scientists; the company's own laboratory records; and complete multi-round materials exploration processes.
The first two are standard practice; every serious post-training effort includes them. The last two are unusual. Laboratory records mean raw operational data from real research environments -- not the idealized textbook version of "mix A with B and heat," but actual research processes with their surrounding context. The multi-round exploration data preserves the full storyline: why this ligand was picked in round one, what the evaluation showed, and how that shaped round two. The model learns exploration trajectories, not isolated facts.
The more radical choice is failed experiments. Deep Principle states explicitly that even failed experiment records went into the training data. In a research context this makes deep sense: failed experiments often carry more information than successful ones -- which path dead-ends, where exactly it broke, how the characterization data exposed the problem. That is precisely the most expensive experience a researcher accumulates. General-model training corpora contain almost exclusively published, successful results (publication bias), so models never learn what trial and error looks like. Feeding in failures turns the tacit knowledge that never makes it into papers into something a model can learn from.
Two training-technique details are also worth noting. First, long-context training was improved to preserve computation, checking, and revision steps inside a reasoning chain -- scientific reasoning runs long, and a model that drops its own intermediate calculations midway will get everything after that wrong. Second, in reinforcement learning, long responses can save progress and be resumed rather than forced out in one pass, which is an engineering prerequisite for exploration loops that run to a hundred rounds or more.
Reading the official scorecard
Every number below is a vendor-reported figure from Deep Principle; independent third-party verification has not caught up yet, so read with that in mind. Even among self-reported numbers, though, some signals are stronger than others.
The headline first. On FSR (FrontierScience-Research), the problem-solving success rate is 38.3%, which the vendor puts above GPT-6-Astra's 33.3%, and 13.9 percentage points above the base GLM-5.3. That 13.9-point jump is more informative than the absolute score: it shows that science-focused post-training genuinely works on a 753B base, rather than being a marketing story. On GPQA-Diamonds, accuracy is 93.0%, up 3 points from the base and, per the vendor, close to Kimi-K3's 93.5%. Note that this is "close to," not "ahead of" -- Kimi-K3 still leads, and the vendor did not paper over it, which is a comparatively honest signal.
The sub-scores cut both ways. Scientific computing averages 76.92% against GPT-6-Astra's 71.79% by the vendor's count; single-step reaction prediction is 90.48% versus 83.33%; experimental design and planning (on a 10-point scale) averages 5.59 against Kimi-K3's 4.57 -- with the caveat, stated by the vendor, that Opus-5.5 and GPT-6-Astra could not be compared fairly on this task because they refused too many prompts, so the comparable field here is narrow. In a multi-round search over metal complexes, the best fifth-round HOMO-LUMO gap averaged 4.318eV, which the vendor puts above GPT-6-Astra's 3.689eV and Kimi-K3's 3.353eV. Getting better the more rounds it screens is exactly what you would expect from training data built on complete multi-round exploration trajectories.
But on NMR spectral interpretation, Hydrogen-1.0 averages 26.67, below GPT-6-Astra's 30.00 on the vendor's own numbers. Spectral interpretation is precisely the task that leans hardest on massive volumes of real spectra, which suggests its science capability is not uniformly distributed and the weak spot is real. The vendor did not publish a table of wins only, and a scorecard that includes a loss is worth more than an all-green marketing chart.
The combined read: if your workflow centers on reaction prediction, experiment planning, or multi-round materials screening, the vendor's numbers show a lead; if your core work is spectral interpretation, do not stop at this one release -- wait for independent evaluation.
Lab Compile: running the experiment plan in a sandbox first
Hydrogen-1.0 also ships with a proprietary mechanism called Lab Compile, which deserves its own section. After the model plans an experiment and refines it into operation instructions, it does not execute directly. Instead, it first "compiles" the plan in a virtual environment, checking whether the equipment supports each action and whether the order of operations is correct. The official blog's example is intercepting an instruction that asks a liquid-handling tool to aspirate when no tip is installed. In a real lab this class of sequencing error means ruined reagents or damaged equipment.
The metaphor is borrowed from software engineering: code has a compiler that catches syntax and type errors before anything runs; experiment instructions never had an equivalent. Lab Compile aims to be that compiler for lab instructions. It pairs directly with the training data described above -- a model that has seen many complete experimental trajectories with their full context knows which orderings are legal and which preconditions must hold. For teams wiring AI into automated laboratory platforms, this checking layer is not a nice-to-have; it is a safety precondition.
Deployment bar and three ways in
Cold water first: self-hosting is extremely demanding. The Machine Heart weekly briefing puts it as requiring large-scale model inference resources; translated into numbers, that is the full BF16 weights at 753.33B parameters and roughly 1.51TB -- merely loading the weights takes a multi-GPU, high-memory cluster, and that is inference, not fine-tuning. Ordinary teams and individual researchers should write off hosting the full version locally.
There are three ways in. First, local deployment with the right quantization: on ModelScope, modelscope.cn/models/deep-principle/Hydrogen-1.0 is the full BF16 weight repository -- deep-principle/Hydrogen-1.0 is the repository name -- and if you run NVIDIA Blackwell hardware (B200/B300) you take the NVFP4 build, while AMD Instinct (MI350X/MI355X) users take the MXFP4 build. Each quantized build is bound to its specified hardware; the two are not interchangeable. The vendor reports deployment has been tested on 8 GPUs with SGLang, which is the official lower-bound reference; if you are setting up yourself, start from SGLang's standard deployment path.
Second, the API route: Deep Principle's own NeuroSpark API is open for access, and you never touch a weight. Third, the product route: the company's AI scientist product, Mira, packages the same capability for people who want results without running infrastructure. The vendor has also announced that Google, AWS, and Microsoft Azure MaaS platforms will carry the model in the future, which will make cloud access easier.
The weights are mirrored on Hugging Face at huggingface.co/deep-principle/Hydrogen-1.0, matching the ModelScope release. The quantized repositories are deep-principle/Hydrogen-1.0-NVFP4 and deep-principle/Hydrogen-1.0-MXFP4 -- pick by hardware.
Finally, the model card's own limitation statement, quoted in spirit: outputs must be reviewed scientifically against task-relevant assumptions, constraints, and evidence, and the official demonstrations combine the model with external tools and specific research environments. It hands you candidate plans and reasoning; the final scientific judgment stays with the researcher. That is the shared boundary of every AI-for-Science tool today.
FAQ
Q1: Can an ordinary team actually run a 753B model? A1: The full version is essentially out of reach -- BF16 weights run about 1.51TB, and local inference needs a multi-GPU, high-memory cluster; the official lower-bound reference is 8 GPUs running SGLang. Teams with B200/B300 or MI350X/MI355X hardware can pick the NVFP4 or MXFP4 build to cut resource needs. Without that hardware, use the NeuroSpark API or Mira instead of fighting a local deployment.
Q2: How is this different from asking a general model science questions? A2: A general model works in answer mode and stops at one output. Hydrogen-1.0 works in process mode -- writing simulation programs, debugging computational anomalies, analyzing reaction pathways, interpreting spectra, planning experiments and refining them into operation instructions, then using each round of evaluation to recommend the next material candidates. Lab Compile additionally checks operation order before execution. The whole design is built around running a research project, not answering a question.
Q3: Can the official benchmark numbers be trusted? A3: Read them at a discount. Every figure is vendor-reported and has no independent third-party verification yet. The more credible signals: the 13.9-point FSR gain over the GLM-5.3 base shows the post-training genuinely worked, and the scorecard includes a loss -- NMR at 26.67 versus GPT-6-Astra's 30.00 -- so it is not cherry-picked. Treat official numbers as a screening tool and test critical tasks in your own real scenarios.
Q4: Does the MIT license mean unrestricted commercial use? A4: MIT is among the most permissive licenses: commercial use, modification, closed-source distribution, and integration into commercial products are all allowed. But mind the boundaries: the weights are MIT, while the training data and full training recipe were not released; the model card requires scientific review of outputs before use. Read the license text yourself before commercializing, and in high-stakes domains like medicine or materials safety, compliance responsibility stays with you.
Q5: Why put failed experiments in the training data? A5: Because failures often carry more information than successes -- which path dead-ends and where it breaks is a researcher's most expensive experience. General-model corpora contain almost only published successes, so models never learn what trial and error looks like. Feeding in failure records and the complete context of multi-round exploration turns the tacit lab knowledge that never reaches a paper into something learnable, which is what lets the model run a real exploration process.
Related reading
- Wuzhen open source: China's new AI open-source hotspot - same-day coverage placing Hydrogen-1.0 on the domestic open-source release timeline.
- Mistral Large 4 open-weight resource guide - the European camp's open-weight route; a useful contrast in open-release strategies.
- SenseNova Loop-9B open-source resource guide - open science models at a very different scale; 9B-class models suit teams without large clusters.