Deep Principle: Hydrogen-1.0 open-weight for science

Deep Principle presents Hydrogen-1.0, a 753 billion parameter MoE with a nominal context of 1 million tokens, specialized from GLM-5.3, and reports an advantage in its own test. The announcement consulted on November 5, 2026 does not state a publication date and does not demonstrate equivalence in production. The decision depends on the workload.

Key facts

• Primary use: support for scientific research questions and problems, with expert review.
• Base: GLM-5.3 with scientific post-training and weights announced for download and via API.
• Benchmark: advantage reported by the vendor on FrontierScience-Research, with no demonstrated independent reproduction.
• Billing: Figure 1 uses output tokens, including reasoning; contractual rates still need confirmation.

Rights to the weights and availability in Brazil remain outside what the announcement covers. Nexforce Router offers billing in BRL with an invoice, but the presence of Hydrogen-1.0 in the catalog is not confirmed.

What changes in the Hydrogen-1.0 open-weight architecture?

Hydrogen-1.0 open-weight specializes GLM-5.3 for scientific tasks through post-training, according to Deep Principle. The vendor describes supervised fine-tuning and reinforcement learning in an MoE model with 753 billion total parameters, a nominal context of 1 million tokens and three weight variants.

The release incorporates data from chemistry, materials and life sciences, plus records and trajectories from the company's own laboratories. The GLM-5.3 announcement states that the base is shared with GLM-5.2. Deep Principle describes the post-training as applied to the entire MoE model, a mixture-of-experts architecture, and combines supervised examples with feedback-driven reinforcement learning, without establishing a chronological order between the two methods. The public Hydrogen-1.0 model card, consulted on November 5, 2026, states a nominal context of 1 million tokens and a maximum output of 128 thousand tokens, with BF16, NVFP4 and MXFP4 variants. The effective limit depends on the runtime, and the model card does not prove the requirements of the buyer's environment. The company does not publish the token proportion of each domain or the volume of trajectories used in the supervised fine-tuning.

Access is not reproduction.

The announcement does not address the license terms for the weights. The terms declared for the weights appear in the distribution repository, and this review did not perform a legal audit. Confirming rights for commercial use remains the buyer's step.

Source: Deep Principle and GLM-5.3 announcements, consulted on November 5, 2026. Art: Nexforce; own diagram, not an independent result.

What does the scientific AI benchmark actually show?

Deep Principle reports that Hydrogen-1.0 surpassed GPT-6-Astra on FrontierScience-Research and improved on GLM-5.3 in GPQA-Diamond. The metrics are different and the vendor ran the evaluations. The results justify an in-house test, but they demonstrate neither overall leadership nor independent reproduction.

In the Figure 1 data, consulted on November 5, 2026, Deep Principle records 38.33% approval for Hydrogen-1.0 and 33.33% for GPT-6-Astra on FrontierScience-Research, both with maximum reasoning. The numbers are the vendor's own results. On GPQA-Diamond, Deep Principle uses accuracy and reports an improvement of Hydrogen-1.0 over GLM-5.3 and a result close to Kimi-K3. Proximity on that test does not demonstrate a tie on every task in the suite. In the internal nuclear magnetic resonance test, the company reports a higher result for GPT-6-Astra at maximum reasoning. The examples show why a single score does not decide usage. To compare versions, prompts and acceptance criteria on equivalent tasks, the team can turn to the model comparison on the same basis. Evaluation of open-weight models should include errors relevant to the domain and expert review of the answers before adoption.

The same file shows GPT-6-Astra above Hydrogen-1.0 on GPQA-Diamond at all three effort levels; at the lowest, the difference falls to 0.5 percentage point. A reported benchmark does not replace a test on the buyer's task.

What does it cost to evaluate and operate Hydrogen-1.0 open-weight?

Figure 1 of the announcement uses standard rates in dollars per output token, including reasoning, and takes the GLM-5.3 price as the reference for Hydrogen-1.0. The calculation excludes input and evaluator. None of that information establishes a contractual rate, total cost or availability on Nexforce Router.

The price reference in Figure 1 does not determine what an organization will pay for Hydrogen-1.0. The buyer should request the current input and output rates, confirm the currency and measure the volume of reasoning on its own tasks before comparing vendors. The calculation needs to include infrastructure, storage, staff, support and evaluation, plus retries and human review when required. Cost per accepted task and the cost of each attempt help separate unit price from spending per useful result. The vendor points to download and API access, but this review does not confirm SLA, region, compatibility or contracting for each customer. Without a published input rate, the estimate of spending per task stays incomplete. Comparison with alternatives requires the same calculation basis. Currency and exchange rates enter the Brazilian buyer's calculation and change over time, so the date of the quote matters.

Note: list rates in dollars, with reasoning tokens included and no batch discounts; input and evaluator stay outside the calculation.

The highest rate does not imply a higher cost per answer on every task. The volume of reasoning tokens changes the bill, and that is what the buyer measures on its own workload.

How to assess integration and routing before contracting?

The announcement offers weight download and an API access route, without describing SLA, region or usage limits. For the buyer, integration needs to be tested with the same tasks as the quality evaluation, because the access path changes cost, latency, control and operational responsibility.

Hydrogen-1.0 open-weight can run on in-house infrastructure or be reached through the announced API route, and each path has distinct cost and governance implications. Running the weights requires hardware compatible with the BF16, NVFP4 or MXFP4 variants, staff to operate the environment and defined update processes. Using the API transfers that operation to the provider and keeps the dependency on availability and price. In either case, the team needs per-call observability, spending limits and a record of the model version in use, because a silent checkpoint change invalidates earlier comparisons. Nexforce Router offers comparison of one prompt across available models, routing rules, spending limits and observability, plus billing in BRL with an invoice. Those functions do not prove that Hydrogen-1.0 is in the catalog or hosted by the Router. Network conditions and rate limits matter as well.

Frequently asked questions about Hydrogen-1.0

Hydrogen-1.0 open-weight combines an existing base and scientific post-training, according to the vendor, Deep Principle. Anyone adopting the model needs to check the weight usage terms, the access conditions and the published results for each intended task before starting any test.

Was the model trained from scratch?

No. Deep Principle states that Hydrogen-1.0 starts from GLM-5.3 and receives scientific post-training. The specialization uses domain data and laboratory records, but the model card does not release with the weights all the data and recipes needed to reproduce the training in full.

Does the license allow any use?

The announcement does not address the license terms for the weights, and the distribution repository publishes the conditions that apply to the weights. Declared terms are not equivalent to unrestricted permission and do not replace a legal audit. The team should confirm the variant, the intended purpose and the attribution obligations before putting the model into production.

Is the GLM-5.3 price the final price?

No. Deep Principle uses the GLM-5.3 output price as a reference in Figure 1, including reasoning. The calculation excludes input and evaluator. The contractual rate and the cost of infrastructure, support and review need to be confirmed for the chosen operation.

References and Further Reading

The primary sources serve different purposes. Deep Principle reports the training, the tests and the cost methodology of Figure 1. The public Hydrogen-1.0 model card on Hugging Face describes variants, nominal context, runtime limits and the declared terms for the weights. Z.ai presents the GLM-5.3 base and its relationship to GLM-5.2.

None of them proves independent reproduction, and the links appear next to the claims they support. The vendor's data file records the scores of 38.33% and 33.33% on FrontierScience-Research at maximum reasoning, consulted on November 5, 2026. To size the choice, the reading on AI cost per task examines spending beyond the rate per token.

When to choose Hydrogen-1.0 open-weight for a task?

Hydrogen-1.0 deserves a test when the task requires scientific knowledge and the team can verify access, quality and cost in the intended environment. The choice for production depends on results on its own workload, on applicable rights and on someone accountable for reviewing answers, not only on the published score.

Selecting Hydrogen-1.0 open-weight should record the tested variant, the versions of the compared models, the prompts and the inference conditions for each run. The published score does not state the error rate of a specific business task or workload, so the team defines acceptance criteria before the test and keeps every output for expert review. The runtime environment needs to be checked before assuming any context limit, because the nominal number does not describe the available capacity. The commercial proposal should separate input, output, reasoning and operating expenses, and also state currency and support. For available alternatives, Nexforce Router can support comparison and spending control. Switching models requires repeating the whole evaluation and preserving the records needed to compare the result. Without those records, the comparison becomes opinion and the test cannot be repeated later.

1. Confirm access and rights for the chosen variant.
2. Measure quality and errors on real tasks.
3. Document infrastructure, security and cost per task.
4. Define support, an accountable owner and a switching plan.

A benchmark is not a decision.