Deep Principle releases Hydrogen 1.0 in INT8 format for download
Image generated by the Inferama Editorial Team
01

What Deep Principle has released

Deep Principle has published an INT8-quantized distribution of Hydrogen-1.0 in a public repository. The checkpoint is a variant of the company’s scientific model, not a new generation of Hydrogen. Its documentation identifies GLM-5.3, from Z.AI, as the base model.

The identifiable public repository revision contains the weights and documentation for this variant. However, the original publication date has not been confirmed: the revision identifies the state of the artifact, but does not by itself establish when it was announced or first became available.

The model is designed for text-based interaction about scientific questions, explanations of derivations, and structured responses to open-ended research problems. Its model card describes demonstrations involving scientific computing, reaction-network exploration, NMR structure analysis, and experimental planning. Those examples combine external tools with specific research environments; on their own, they do not establish general performance or verified scientific results.

The release offers a deployment option with a smaller weight footprint than the BF16 distribution, in exchange for compatibility requirements and possible effects from quantization. Lower memory use does not guarantee greater speed or accuracy. The model card explicitly says that this variant is not expected to be faster than BF16.

02

Access, size and availability

The repository provides the INT8 weights and their configuration files for download and use in a compatible environment. The checkpoint is approximately 793 GB, or 739 GiB, compared with about 1.5 TB for the BF16 variant listed in the model card. The smaller footprint may matter when planning storage and memory, although deployment still requires substantial resources.

Publishing the weights does not confirm a public hosted API or a cloud usage rate. Anyone wanting to try the model needs their own infrastructure or a provider that supports this checkpoint; compute and storage costs depend on that environment. The model card does not publish input or output prices for a hosted API.

The documentation states a context length of up to 1 million tokens and a maximum output length of up to 128,000 tokens, but notes that effective limits depend on the inference framework, service configuration, and available resources. The example launch command sets the context to 262,144 tokens, below the model’s stated maximum. These advertised maxima should not be confused with the limit available in each deployment.

The weight files in the identified repository revision establish that a downloadable distribution exists there. They do not, on their own, establish that a managed service is available, that every user can deploy it on their own hardware, or that all providers offer the same context limit.

What the availability details mean

The model card distinguishes the downloadable checkpoint from the limits determined by each deployment.

DetailWhat it supportsWhat it does not confirm
INT8 weights in the repositoryA weight distribution is available in the identified revision.A hosted API or the original publication date.
Stated context length of up to 1 million tokensThe model and its card describe this as an intended maximum capacity.That every runtime or server can enable it.
Deployment example with 262,144 tokensThat specific configuration sets a lower limit.That every installation must use that limit.
No API price publishedThe model card does not confirm a hosted rate.That using your own infrastructure is free.
03

What INT8 quantization changes

INT8 describes how certain model weights are stored as 8-bit integers. For this distribution, Deep Principle says it quantizes the routed experts in layers 3–77 and the dense MLP networks in layers 0–2, using symmetric weights and one scale per group of 128 weights. Activations and several shared components, including attention, the router, and embeddings, remain in BF16.

The partial scope matters: INT8 does not mean every model component uses 8-bit integers. The model card describes weight-only quantization applied directly to the BF16 weights without calibration data. It also identifies the checkpoint format as compressed-tensors pack-quantized, so a compatible runtime must support that format and its integer-weight path for MoE models.

The practical benefit described by the provider is a smaller weight footprint than BF16. The trade-off is that users need to retain the configuration files and use an inference engine that supports the combination of INT8 and BF16 components. The format is not tied to a particular GPU brand, but that portability does not show that every hardware setup has been tested.

The documentation warns that quantization may affect outputs and recommends validating the model for the intended task. It publishes no numerical accuracy measurement specific to this INT8 variant, so it does not quantify how quality compares with BF16.

04

Requirements and scope of testing

Deep Principle says it tested the checkpoint with SGLang 0.5.20 on eight NVIDIA B300 GPUs. The model card limits this claim to that setup: other hardware and runtimes are not listed as tested. The result is a configuration reference, not a guarantee that the installation will work on every server.

The launch example requires disabling shared-expert fusion. The stated reason is that shared experts use BF16 while routed experts use INT8, so the runtime cannot fuse those weights into a single expert tensor. According to the documentation, omitting this option prevents SGLang from placing the shared-expert weights correctly.

The deployment should also keep the checkpoint’s configuration files together, including config.json, the tokenizer, chat template, generation_config.json, and weight index. These files let the runtime use the quantization settings and generation conventions associated with this distribution. Substituting files from another variant may change the effective configuration or prevent correct loading.

The model card also advises users to review remote code before running it. Format compatibility, available memory, and context configuration are separate conditions: meeting one does not guarantee the others. In particular, raising the context limit above the example requires compatible runtime support and enough memory.

Checks before deployment

Use these steps to assess whether an environment may be able to run the checkpoint; they do not replace operational testing.

  1. 01Check that the runtime supports the compressed-tensors pack-quantized format and an integer-weight path compatible with MoE models.
  2. 02Verify that available memory can accommodate the weights, requested context, and the service’s additional resource needs.
  3. 03Keep the weights, config.json, tokenizer, chat template, generation_config.json, and weight index from the same distribution together.
  4. 04Configure shared-expert handling as required by the checkpoint; the SGLang example includes --disable-shared-experts-fusion.
  5. 05Start with a context limit appropriate for the available memory and runtime, and increase it only if the environment supports doing so.
  6. 06Test responses on the actual tasks and independently verify scientific conclusions.
05

What it is for—and what it does not establish

Hydrogen-1.0 is intended for educational exploration and support with scientific tasks: explaining equations, working through derivations under explicit assumptions, and preparing structured drafts for expert review. The model card does not present its answers as confirmed experimental results. A detailed or confident explanation does not establish that it is correct.

The model may produce errors in facts, calculations, equations, references, or explanations. Its performance may also vary by discipline, language, task format, and input length. A long context does not guarantee that the model will retrieve, understand, or correctly use every detail provided.

The model card says its evaluation focuses on scientific multiple-choice questions and open-ended research problems. It gives no numerical results specific to this INT8 variant. Demonstrations for the original model should not automatically be transferred to this distribution or treated as independent replications.

To evaluate the model for a particular task, prepare representative examples, check calculations independently, and verify citations and claims against primary literature. For applications involving health, safety, or other consequential decisions, the answers are not a sufficient basis for a decision without competent review.

06

Licensing and included materials

The quantized Hydrogen-1.0 INT8 weights are distributed under the MIT License included in the repository. The license requires retaining the copyright and permission notices when using or redistributing the weights. The licensing document attributes the GLM-5.3 base model to Z.AI and clarifies that Hydrogen’s license does not change the license for a separately obtained original GLM-5.3 checkpoint.

Documentation and original examples written by Deep Principle have a different licensing scope and are identified as Apache License 2.0. Third-party files, such as configurations, tokenizers, templates, or code, retain their respective licenses and notices. The MIT License for the weights does not automatically make every related component MIT-licensed.

The repository does not include complete training datasets or complete training recipes. The licenses for the published files do not grant access to or redistribution rights for training materials that are not part of the distribution.

The verifiable release is an INT8 variant available for download, with smaller weights than BF16 and specific runtime requirements. The original publication date and the existence of a hosted service remain unconfirmed; no published numerical evaluation measures the effect of this quantization on quality.

07

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction