Hydrogen-1.0

Hydrogen-1.0 is an open-weight scientific research large language model developed by Deep Principle from Z.AI's GLM-5.3. Announced on November 2, 2026, it specializes in scientific problem solving through additional training on scientific data and laboratory work. The release includes BF16 weights and separate quantized checkpoints for NVIDIA and AMD hardware.^[1][2][3]

The model is intended to help researchers reason through questions, explain derivations and prepare research proposals for review. Its model card distinguishes those outputs from verified experimental findings.^[1]

Model and training

These are model-level specifications. The supported input and output budgets also depend on the serving framework, memory and deployment settings.^[1]

Deep Principle describes supervised fine-tuning followed by reinforcement learning. Its data includes chemistry, materials science and life sciences knowledge, task feedback from scientists, laboratory records, and sequences of research decisions and observations. The training infrastructure builds on Megatron and SGLang across CUDA and ROCm. The company releases weights without the complete training datasets or training recipes.^[1][2]

Scientific evaluations

Public benchmarks

Deep Principle's co-founder and chief technology officer Chenru Duan reported the following scores in the launch announcement. They are publisher-reported results, rather than an independent evaluation of the checkpoint.^[3]

FrontierScience is an OpenAI benchmark with separate Olympiad and Research tracks. The published Research gold set contains 60 original subtasks written by PhD scientists in physics, chemistry and biology, graded with a 10-point rubric. Those tasks test bounded scientific reasoning problems. They do not reproduce an entire research project, which also involves choosing questions, conducting experiments and assessing ambiguous evidence.^[6]

GPQA uses expert-written multiple-choice questions in biology, physics and chemistry. Its Diamond subset selects questions with stronger expert agreement and low non-expert success. It tests a different task format from research planning or laboratory execution, so its accuracy percentage and a FrontierScience pass rate describe different capabilities.^[7]

Internal tasks and demonstrations

Deep Principle also compares Hydrogen-1.0 with GPT-6 Astra on internal scientific computing, forward reaction prediction and NMR structure tasks. Against Astra at maximum thinking, it reports leads of 5.13 and 4.76 percentage points on the first two tasks, and a 3.33-point deficit on NMR. These task-specific results should retain their evaluation context.^[2]

The release demonstrates a polymer molecular-dynamics workflow, reaction-network exploration and NMR analysis. Its laboratory-protocol demonstration is explicitly a scripted virtual replay with capability and prerequisite checks, without a live device connection.^[2]

Checkpoints and deployment

The BF16 documentation instructs operators to retain generation_config.json, which contains end-of-turn settings needed for stopping. The matching tokenizer and chat template also belong with the checkpoint. A compatible API server can expose the model under the identifier hydrogen-1.0; the model card's client example calls an existing service rather than downloading weights or creating a hosted endpoint.^[1]

NVIDIA NVFP4 version

The NVFP4 checkpoint uses NVIDIA ModelOpt quantization. Its routed experts use NVFP4, while the dense MLPs in layers 0-2 use FP8. Shared experts, attention, routing, embeddings, the language-model head and the multi-token prediction layer remain BF16. The export also provides FP8 KV-cache settings.^[4]

The published SGLang example selects modelopt_fp4, FP8 E4M3 KV cache, and tensor parallelism of eight. It explicitly caps the deployed context at 262,144 tokens. This example therefore demonstrates a smaller configured context than the model's stated 1-million-token capacity. Increasing that limit requires compatible runtime support and sufficient memory.^[4]

AMD MXFP4 version

The MXFP4 checkpoint uses AMD Quark. Its dense MLPs, shared and routed experts, and multi-token prediction MLPs use MXFP4 with dynamic activations. Attention, the router, embeddings, the language-model head and eh_proj remain BF16. Weight and activation groups contain 32 elements with E8M0 scales.^[5]

Deep Principle reports an ATOM deployment using tensor parallelism of four and FP8 KV cache, with speculative decoding disabled. In that setup, 128 concurrent long reasoning requests exhausted the cache pool and caused eviction alongside a sharp answer-quality decline. At 48 concurrent requests, utilization stayed below 75% and quality was normal. These observations describe that test configuration, rather than general concurrency limits for all hardware.^[5]

Access, licensing and limitations

The release announcement lists weight downloads through Hugging Face and ModelScope, API access through NeuroSpark, and use within Deep Principle's Mira AI Scientist. The cost figures in its benchmark plots use GLM-5.3 rates for Hydrogen-1.0; the plots are evaluation cost estimates.^[2]

Deep Principle labels the Hydrogen-1.0 weights MIT-licensed and its independently authored code and documentation Apache-2.0. Its model cards state that third-party material keeps its respective license. The BF16 repository also requires users to review access conditions through Hugging Face.^[1][4][5]

The model card warns about incorrect scientific facts, calculations, references and explanations, as well as performance variation across disciplines and input lengths. Quantized versions can produce different answers from BF16. Research conclusions therefore require checks against primary literature, calculations or experiments, even when a response is detailed and confident.^[1]

Continue reading

• GLM-5.3GLM-5.3 is a text-input and text-output large language model from Z.ai, available through hosted services and downloadable checkpoints.Mentioned 3 times above
• Large Language ModelA large language model (LLM) is an artificial intelligence system built on a transformer neural network with billions to trillions of parameters, trained on massive text corpora to predict the next token in a…Linked in the introduction
• SGLangSGLang (short for Structured Generation Language) is a high-performance, open-source serving framework for large language models and multimodal models, originally developed at UC Berkeley's Sky Computing Lab…Mentioned 4 times above
• GPT-6 AstraGPT-6 Astra is an OpenAI model that the company began rolling out on September 3, 2026, describing it as "the world's most intelligent and aligned model" and as the successor to the GPT-5.6 family.Mentioned twice above
• Evo 2 (genomic model)Evo 2 is a genomic foundation model built by the Arc Institute together with NVIDIA, Stanford University, and collaborators, and first released in February 2025.Also in AI for Science
• Boltz-2Boltz-2 is an open-source biomolecular foundation model that jointly predicts the 3D structure of biological complexes and the binding affinity between small molecules and proteins.Also in AI for Science

More to read

• Boltz
• Llama 3
• DeepSeek V3
• Qwen3
• Llama 3.1
• Kimi K3

Topics:AI ModelsAI for ScienceLarge Language ModelsOpen Source AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Deep Principle. Hydrogen-1.0 model card. Checked November 4, 2026.
2. ^1 ^2 ^3 ^4 ^5Deep Principle. Introducing Hydrogen-1.0. Technical release. Checked November 4, 2026.
3. ^1 ^2Chenru Duan. Hydrogen-1.0 launch announcement. November 2, 2026.
4. ^1 ^2 ^3 ^4Deep Principle. Hydrogen-1.0-NVFP4 model card. Checked November 4, 2026.
5. ^1 ^2 ^3 ^4Deep Principle. Hydrogen-1.0-MXFP4 model card. Checked November 4, 2026.
6. ^OpenAI. Evaluating AI's ability to perform scientific research tasks. December 16, 2025.
7. ^David Rein et al. GPQA: A Graduate-Level Google-Proof Q&A Benchmark. arXiv:2311.12022, November 20, 2023.