Hydrogen-1.0: Scientific Tasks, API Access and Deployment

Deep Principle applies supervised fine-tuning and reinforcement learning to GLM-5.3 using scientific knowledge, research scenarios, laboratory records and discovery trajectories. Hydrogen is a separate post-trained checkpoint. Its scientific specialization is the reason to compare it with the general-purpose base on your own research questions.

The model card lists 753 billion parameters, a one-million-token context and up to 128K output tokens. A deployment’s available memory, request limits and output budget determine how much of that capacity a task can use. Start with a bounded document set and a question whose answer you can verify.

The release links to NeuroSpark for API access and to Deep Principle’s Mira research platform. Use the service’s account documentation for its endpoint, model identifier, quotas and current billing terms. The publisher’s public release and model card do not provide a NeuroSpark token-price table.

Download the weights from the publisher’s Hugging Face or ModelScope repositories. Hugging Face gates the BF16 download and some raw repository files behind its access conditions; the rendered model cards describe the formats and deployment recipes. NVIDIA NVFP4 and AMD MXFP4 versions have separate repositories and deployment instructions. Select a format for the actual accelerator and serving engine rather than substituting one quantized download into another format’s command.

Hydrogen-1.0 has separate BF16, NVFP4 and MXFP4 artifacts. Use the runtime and parallel layout documented for your chosen checkpoint. The publisher-tested configurations below do not establish minimum memory or interchangeable GPU support.

The documented NVFP4 launch sets --context-length 262144: input and generated output must fit a served context of 262,144 tokens, below the model card’s stated 1M. Increasing that limit requires compatible runtime support and sufficient memory. Do not size a document workflow against the model-level maximum while keeping this recipe unchanged.

For the BF16 recipe, preserve generation_config.json and use the documented glm45 reasoning parser and glm47 tool-call parser. The card warns that a missing generation configuration can prevent output from stopping at the end of an answer. Its launch recipe enables trust_remote_code, so review the checkpoint code before starting it in your serving environment.

After the server is running, its served model name is hydrogen-1.0. The request below targets your own SGLang endpoint; it is not a NeuroSpark request. Replace the base URL and authentication setup with those of your server. A 4,096-token cap is a first integration budget and can truncate reasoning on a complex question.

Obtain the hosted service’s current quote before forecasting API spend. The launch article’s benchmark-cost chart uses GLM-5.3 output-token pricing as a proxy, including reasoning tokens and excluding input and evaluation-judge charges. It is not a Hydrogen API tariff or a full experiment bill.

For self-hosting, add reserved accelerator time, storage and the external computation your workflow performs. If an assumed eight-GPU server costs $24/hour and is reserved for 30 minutes, that server time costs $12 before storage or simulation tools. This is an illustrative rental assumption; the deployment recipe does not establish that hourly rate or how many tasks it can finish.

Count model retries, invalid scientific candidates and failed tool executions in the total. Divide by accepted results after expert review. A molecular simulation can use compute beyond model inference, and a proposed experiment can require human and laboratory resources even when generating the plan is inexpensive.

Deep Principle reports improvements over GLM-5.3 on scientific evaluations and comparisons with other frontier models. Its internal task results vary: Hydrogen improves scientific-computing and forward-reaction results in the reported comparison with GPT-6 Astra, while the NMR task favors Astra. Select an evaluation close to your discipline rather than treating the release as a universal scientific ranking.

The publisher demonstrates a polyethylene molecular-dynamics workflow that corrects a topology error and analyzes a 40-nanosecond trajectory. Another example proposes a reaction network, and a Suzuki protocol example repairs a missing pipette-tip change during virtual robot rehearsal. These illustrate tool-assisted workflows in the publisher’s environment; the robot rehearsal did not connect to a real instrument.

For a first evaluation, keep a known calculation, expected units and a source you can inspect. Ask the model to state assumptions, run the code under your own checks and compare the final result with that reference. A plausible narrative or a completed virtual rehearsal does not establish an experimental finding.

The Hydrogen BF16 repository labels its weights MIT. Retain the copyright and permission notices supplied with the checkpoint. The model card separately labels independently authored Deep Principle code and documentation Apache-2.0, with third-party files retaining their own terms. Preserve the notices for the artifacts you redistribute.

The GLM-5.3 base has its own custom license. For an offering built on Hydrogen, inspect the Hydrogen repository’s license, licensing-scope and attribution files together with the inherited materials. The weight label alone does not describe every file in a deployment.