Deep Principle Launches Hydrogen-1.0, an Open-Weight Scientific Model Built on GLM-5.3

Hydrogen-1.0 is a 753B-parameter open-weight mixture-of-experts model released by Deep Principle on Nov. 2, 2026, created by post-training Z.ai's GLM-5.3 on chemistry, materials science, life sciences, laboratory records and discovery trajectories.

Deep Principle released Hydrogen-1.0 on Nov. 2, 2026, positioning the 753B-parameter mixture-of-experts model as its first open-weight system built specifically for scientific research. The model arrives under the MIT license on Hugging Face behind a gated download and on ModelScope, with additional access through the NeuroSpark API and the Mira AI Scientist application. In its release post, Deep Principle said the model was built by post-training GLM-5.3, the open-weight model from Z.ai, on chemistry, materials science and life sciences data combined with records and agent trajectories from the company's own laboratories.

The release inserts a domain-specialized entrant into a fast-moving open science model market. Hydrogen-1.0 inherits roughly 40B active parameters per token from its GLM-5.3 base, supports a 1M-token context window and a 128K maximum output length, and is testable on eight GPUs using SGLang, according to Deep Principle. Quantized versions target NVIDIA Blackwell and AMD Instinct hardware. The company self-reports a score of 38.3% on its FrontierScience-Research benchmark, compared with 33.3% for GPT-6-Astra, and 93.0% on GPQA-Diamonds, a result Deep Principle says roughly matches Kimi-K3.

What is Hydrogen-1.0 and how does it relate to GLM-5.3?

Hydrogen-1.0 is a post-trained derivative rather than a newly pretrained base model. Deep Principle took GLM-5.3, Z.ai's general-purpose open-weight model, and ran an additional training phase on scientific domain data, laboratory records and discovery trajectories. The distinction matters because post-training preserves the base model's architecture and broad capabilities while reshaping its behavior toward specialized tasks, in this case scientific problem-solving and research assistance.

The Hugging Face model card describes Hydrogen-1.0 as a scientific problem-solving and research assistance model developed by Deep Principle based on GLM-5.3, listing 753B parameters, a 1M-token context length and the MIT license. The mixture-of-experts design keeps compute costs manageable at inference time: only about 40B parameters are active for any given token, a property inherited from the GLM-5.3 base. That efficiency profile is what makes the model practical to run on eight GPUs with SGLang rather than requiring a large cluster.

The choice of GLM-5.3 as a base is itself a strategic signal. Rather than pretraining a model from scratch, a process that requires thousands of GPUs and hundreds of millions of dollars, Deep Principle adapted an existing open-weight model at a fraction of that cost. Z.ai's GLM line has been widely used for derivative work, and the MIT license on Hydrogen-1.0 extends the same reuse logic one step further down the stack, allowing downstream developers to modify and commercialize the scientific model.

How did Deep Principle build and release Hydrogen-1.0?

Deep Principle's stated differentiator is data, not architecture. The company says it post-trained the model on domain data from chemistry, materials science and life sciences, together with records and agent trajectories from its own laboratories. Those trajectories, logged sequences of actions an agent takes while pursuing a research goal, are the kind of asset that is difficult for outside labs to replicate because they are generated inside Deep Principle's experimental infrastructure.

The release follows a specific sequence of availability channels:

1. Gated download on Hugging Face under the MIT license.
2. Mirrored listing on ModelScope for users in markets where Hugging Face access is restricted.
3. Managed access through the NeuroSpark API and the Mira AI Scientist application.
4. Quantized builds for NVIDIA Blackwell and AMD Instinct GPUs.
5. Public release of weights without accompanying training datasets or full training recipes.

The gap between open weights and closed data is the most consequential detail in the release. Deep Principle makes the model's parameters freely usable and modifiable under MIT terms, which permits commercial use with attribution, but it withheld the training data and recipes that would allow others to reproduce or extend the scientific post-training. For research institutions evaluating the model, that means the weights are inspectable while the claimed data advantage is not.

What do the benchmark claims show?

Deep Principle reports two headline numbers. On FrontierScience-Research, an evaluation Deep Principle uses to test scientific research capability, Hydrogen-1.0 scores 38.3% against 33.3% for GPT-6-Astra, a five-point edge. On GPQA-Diamonds, a graduate-level reasoning benchmark drawn from science questions, the model scores 93.0%, which Deep Principle says roughly matches Kimi-K3.

Both figures are self-reported, and they come with different evidentiary weight. GPQA-Diamonds is an externally defined benchmark with published questions, which makes the 93.0% result independently checkable in principle. FrontierScience-Research, by contrast, is Deep Principle's own evaluation, so the comparison against GPT-6-Astra rests on methodology the company has not fully disclosed. Independent verification has not yet been published.

We're introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3 (@Zai_org) on our internal scientific data, lab records, and discovery trajectories. It scores 38.3% on FrontierScience-Research vs GPT-6-Astra's 33.3% and scores 93.0% on GPQA-Diamonds.Chenru Duan, co-founder, Deep Principle

Duan's framing ties the benchmark results directly to the training approach: internal scientific data, lab records and discovery trajectories. The claim that domain post-training can lift a general model's science performance while preserving reasoning quality, which the GPQA-Diamonds result roughly matching Kimi-K3 is offered as evidence of, is the core empirical assertion in the release. It is also the claim most likely to attract replication attempts from other labs.

How does Hydrogen-1.0 compare with LongCat-2.5 and other models?

Hydrogen-1.0 entered a market that includes Meituan's LongCat-2.5, a separately developed 1.6T-parameter multimodal agent model with about 48B active parameters per token, released the same week. The two systems are not direct substitutes: Hydrogen-1.0 is an open-weight scientific research model, while LongCat-2.5 is a general-purpose agent platform sold through the LongCat API. But they compete for the same researcher and enterprise budgets, which makes their specs and pricing directly comparable.

Hydrogen-1.0 versus LongCat-2.5 and related models

Model	Developer	Parameters (total / active)	Context window	Access	Notable claim or pricing
Hydrogen-1.0	Deep Principle	753B / ~40B	1M tokens	MIT open weights on Hugging Face and ModelScope	38.3% on FrontierScience-Research (self-reported)
LongCat-2.5	Meituan	1.6T / ~48B	1M tokens	LongCat API	$0.30 per 1M input tokens; $1.20 per 1M output tokens
GLM-5.3	Z.ai	Base model for Hydrogen-1.0	Not stated	Open weights	Foundation for Hydrogen-1.0 post-training
GPT-6-Astra	—	Not stated	Not stated	Not stated	33.3% on FrontierScience-Research (per Deep Principle)

The LongCat API documentation lists pay-as-you-go pricing at $0.30 per 1M uncached input tokens and $1.20 per 1M output tokens for LongCat-2.5. That price point is notable because it gives the open science segment a commercial reference: researchers who cannot run Hydrogen-1.0's 753B parameters locally can instead rent comparable agent capacity per token, while institutions with GPU capacity can self-host the MIT-licensed weights and avoid per-token fees entirely.

What does the release mean for the open science model market?

Deep Principle's strategy treats proprietary laboratory data as the defensible asset in open-weight AI. By releasing Hydrogen-1.0's weights under MIT while keeping training data private, the company invites adoption of its model while preserving the data moat that would be required to reproduce it. That pattern, open weights with closed data, is becoming the standard shape of domain-specialized model releases, and it puts pressure on generalist labs that lack equivalent experimental infrastructure.

For research institutions, the practical implications are concrete. Hydrogen-1.0 runs on eight GPUs with SGLang, and quantized versions for NVIDIA Blackwell and AMD Instinct reduce hardware requirements further. A university lab can therefore deploy the model in-house, a significant advantage for institutions handling sensitive or proprietary experimental data that cannot be sent to external APIs. The gated Hugging Face release adds a registration step but does not restrict what can be done with the weights once downloaded.

Commercial stakeholders face a different calculation. The $0.30/$1.20 per 1M token pricing attached to LongCat-2.5 establishes an expectation for agent-grade scientific models, and Deep Principle's dual distribution, open weights for self-hosting plus the NeuroSpark API and the Mira AI Scientist app for managed use, lets the company serve both cost-sensitive labs and teams that prefer a hosted workflow.

The competitive context includes not just Meituan but the broader ecosystem of open science efforts. The presence of multiple capable entrants, Deep Principle with scientific post-training and Meituan with agent infrastructure, suggests the open-weight segment is maturing beyond chat assistants into domain-specific tooling. For publishers and enterprise buyers tracking the category, the meaningful question is no longer whether open models can reach frontier capability, but which organizations control the data and evaluation pipelines that make those capabilities useful in practice.

What are the limitations and open questions?

The release leaves several questions unresolved. The benchmark results are self-reported and have not been audited by a third party, and FrontierScience-Research is an internal evaluation whose methodology has not been published in full. The training data and recipes are undisclosed, so claims about the value of laboratory records and discovery trajectories cannot be independently tested. And while the MIT license permits broad use, the gated download means Deep Principle retains visibility into who is accessing the model, a governance choice that some open-source advocates may scrutinize.

There is also the question of what open means in this release. The MIT license is permissive, but the gated download, the withheld training data and the absence of a published data card mean the release does not meet every definition of open science. Deep Principle's terminology, open-weight, is more precise than open-source, and the distinction matters for reproducibility-focused institutions that want to audit how the model was built, not just what it can do.

How will the scientific community evaluate Hydrogen-1.0?

The most likely verification path is a third-party rerun of GPQA-Diamonds, since the benchmark's questions are public and the scoring procedure is established. A similar check on FrontierScience-Research is harder because the benchmark is internal; external labs can test the model on their own held-out science problems, but they cannot replicate Deep Principle's exact comparison with GPT-6-Astra without access to the evaluation harness.

Replication of the training approach is a separate and longer-term effort. Because the data and recipes are withheld, no outside lab can currently reproduce Hydrogen-1.0's post-training from GLM-5.3. That makes the release's central empirical claim, that laboratory records and discovery trajectories materially improve scientific reasoning, an assertion that will be tested only if Deep Principle later opens its data or if competitors achieve comparable results through their own proprietary data.

What comes next for Deep Principle and open science models?

Deep Principle frames Hydrogen-1.0 as an introduction rather than an endpoint. The company's release post describes the model as its first open-weight scientific model, and the infrastructure built for it, including lab records, discovery trajectories and benchmark tooling, is reusable for future iterations. If the self-reported results hold up under external evaluation, the post-training approach could become a template for other organizations with proprietary scientific data but no frontier-scale pretraining capability.

The broader trajectory points toward competition defined by data ownership and agent tooling rather than raw parameter counts. Meituan's LongCat-2.5 brings agent infrastructure and aggressive per-token pricing; Deep Principle brings domain data and open weights. The near-term test for both is whether scientific users can convert benchmark scores into reproducible laboratory outcomes, the same standard that will determine whether Hydrogen-1.0's five-point edge over GPT-6-Astra matters outside the company's own evaluation suite.