Deep Principle released Hydrogen-1.0 on November 2, 2026, as an open-weight AI model for scientific problem-solving. Built by post-training Z.ai’s GLM-5.3, it targets chemistry, materials science, and life sciences. The release includes MIT-licensed BF16 weights, a stated one-million-token context window, and support for up to 128,000 output tokens, according to its Hugging Face model card. Neural Brew’s dated coverage identifies November 2 as the announcement date.
Researchers get a model they can download, deploy, and adapt, with separate quantized versions for NVIDIA and AMD accelerators. But an important boundary remains: the weights are available, while the full training datasets and recipes are not.
That distinction shapes what researchers can reproduce. A team can evaluate the released checkpoint on its own scientific problems without receiving enough information to recreate the training process that produced it.
What the Open-Weight Release Includes
The model card lists the following specifications:
| Item | Published specification |
|---|---|
| Base model | GLM-5.3 |
| Main release | BF16 weights |
| Model license | MIT |
| Context window | 1 million tokens |
| Maximum output | 128,000 tokens |
| NVIDIA quantized version | NVFP4 for Blackwell B200 and B300 |
| AMD quantized version | MXFP4 for Instinct MI350X and MI355X |
| Reported deployment test | SGLang across eight GPUs |
BF16 uses a 16-bit numerical representation for weights. The NVFP4 and MXFP4 releases use hardware-specific four-bit floating-point formats, reducing weight-storage requirements. They are separate artifacts intended for the accelerators listed above, so researchers should choose a build that matches their hardware and supported runtime.
The MIT license is permissive, including for modification and commercial use, subject to its terms. Download access has a separate condition: the Hugging Face page asks users to agree to share their account information before accessing the model.
A Million Tokens Is a Capacity, Not a Reliability Score
A large context window could help researchers supply laboratory records, papers, protocols, and previous results in one request. The practical benefit is less manual selection and fragmentation of the material a model needs to consider.
The model card qualifies both context and output limits, however: available budgets depend on the inference framework, server configuration, and runtime resources. Deploying the weights does not automatically make the advertised maximum available.
Long-context capacity also does not guarantee reliable retrieval or use of every detail. My view is that a lab should test whether Hydrogen-1.0 can preserve decisive constraints across long records before building a workflow around the headline window. Missing a reagent concentration or an experimental exclusion can matter more than producing a fluent summary.
Training Disclosures Explain the Approach, Not the Recipe
Deep Principle’s announcement describes post-training on scientific knowledge, laboratory records, and discovery trajectories from its own laboratories. The domains include chemistry, materials science, and life sciences. Neural Brew reports that the process used supervised fine-tuning followed by reinforcement learning.
Supervised fine-tuning teaches a model from examples of desired behavior. Reinforcement learning adjusts behavior using feedback or rewards. Those broad method names do not reveal the examples, reward design, filtering decisions, or training settings needed to reconstruct this particular model.
The laboratory trajectories are the more distinctive part of the disclosed approach. Deep Principle describes records connecting hypotheses, actions, observations, and revisions, including failed attempts and unexpected results. Such sequences could help teach how evidence changes a research plan, beyond answering isolated scientific questions.
That is a plausible rationale, not proof that the training produced reliable research judgment. The model card explicitly says that “full datasets and recipes are not released with the weights.”
Open weights therefore provide a useful but partial form of reproducibility. Researchers can inspect and evaluate the resulting model, while the training-data composition and complete post-training procedure remain unavailable for independent reconstruction.
What the Company’s Benchmarks Actually Show
Deep Principle reports that Hydrogen-1.0 beats GPT-6 Astra on FrontierScience-Research and improves over its GLM-5.3 starting point on GPQA Diamond, roughly matching Kimi K3. These are claims from the company’s own evaluation.
Neural Brew attributes scores of 38.3% on FrontierScience-Research, versus 33.3% for GPT-6 Astra, and 93.0% on GPQA Diamond to Deep Principle’s Chenru Duan. Neither source provides independent replication, and Neural Brew reports that the internal scientific evaluation sets are not public.
The company also publishes comparisons on three research tasks:
| Internal task | Company-reported comparison with GPT-6 Astra at maximum thinking |
|---|---|
| Scientific computing | Hydrogen-1.0 ahead by 5.13 percentage points |
| One-step forward reaction prediction | Hydrogen-1.0 ahead by 4.76 percentage points |
| NMR structure elucidation | GPT-6 Astra ahead by 3.33 percentage points |
These results support a narrower interpretation than an across-the-board scientific lead. In the company’s tests, Hydrogen-1.0 performs better on two scenarios and worse on another. Deep Principle also warns that differences in absolute scores across tasks reflect the difficulty of the internal cases, so they should not be read as direct comparisons of capability between disciplines.
I’d use these numbers to choose evaluation tasks, not to select a model outright. A lab working on spectral interpretation has a different decision to make from one automating simulation scripts, and the published comparisons already show why that distinction matters.
The Cost Charts Use Substitute Pricing
Deep Principle’s performance-versus-cost charts use GLM-5.3 pricing for Hydrogen-1.0. That makes their economics an estimate based on another model’s API rates, not a measurement of running the released weights.
The announcement also uses different accounting boundaries:
- The scientific-reasoning chart counts output-token costs, including reasoning, but excludes input and judge costs.
- The internal-task chart counts input and output costs, including reasoning and recorded cache discounts.
A self-hosted deployment adds different variables: accelerator rental or purchase, utilization, serving overhead, and the memory demands of long requests. The published charts can illustrate the company’s comparison under its stated assumptions. They cannot establish a lab’s cost per useful, verified result.
Deployment Has Consequential Configuration Details
The model card says Hydrogen-1.0 was tested with SGLang on eight GPUs. That is a documented configuration, not a complete minimum-hardware specification.
Deep Principle provides this launch example:
python -m sglang.launch_server \
--model-path deep-principle/Hydrogen-1.0 \
--served-model-name hydrogen-1.0 \
--tp 8 --trust-remote-code \
--reasoning-parser glm45 --tool-call-parser glm47 \
--host 0.0.0.0 --port 8001
The command loads the main model repository, distributes execution through eight-way tensor parallelism, and exposes a service on port 8001. It also selects the reasoning and tool-call parsers shown in the deployment documentation. This is a published example, not a configuration tested for this article.
Two flags deserve attention before copying it. --trust-remote-code permits repository-provided code to execute, which warrants code review and an appropriately isolated environment. Binding to makes the service listen on all network interfaces; access controls and network restrictions should be in place before exposing it beyond a trusted environment.
Frequently Asked Questions
3 questions
1Is Hydrogen-1.0 Fully Open Source?
Its weights are released under the MIT license, but the full training datasets and recipes are unavailable. “Open-weight” is the more precise description of the release.
