Parameters
753.3B total / 43.0B active
MoE: total / active
Architecture
MoE with IndexShare sparse attention (DSA) + MTP layer
Released
02.11.2026
License
MIT License
Input Modalities
Output Modalities
Context (native)
1,000,000 tokens
Context (extended)
—
About
Hydrogen-1.0 (deep-principle/Hydrogen-1.0) is a scientific problem-solving and research assistance model developed by Deep Principle, built as a full-model MoE post-training of GLM-5.3 (753B total / 43B active, DSA + MTP) on domain data from chemistry, materials science and life sciences, together with laboratory records and discovery trajectories from Deep Principle's own labs. "Through text interaction, it is intended to assist with scientific questions, explain derivations, and organize responses to open-ended research problems. Outputs should be assessed against the assumptions, constraints, and evidence relevant to each task."
Versions — BF16 unquantized weights (this repo); NVFP4 (deep-principle/Hydrogen-1.0-NVFP4, NVIDIA Blackwell B200/B300) and MXFP4 (deep-principle/Hydrogen-1.0-MXFP4, AMD Instinct MI350X/MI355X) are separate repos. Context length 1M tokens, maximum output length 128K tokens (actual budgets depend on inference framework and runtime).
Training — pipeline combines supervised fine-tuning (SFT) and reinforcement learning (RL) on domain knowledge, research scenarios with scientist feedback, laboratory records, and discovery trajectories; infrastructure on Megatron + SGLang across CUDA and ROCm. Full datasets and training recipes are not released.
Evaluation — outperforms GPT-6-Astra on FrontierScience-Research (FSR) and improves over GLM-5.3 on GPQA-Diamond, roughly matching Kimi-K3 (per Deep Principle's evaluations); internal evals cover scientific computing, reaction prediction, NMR structure elucidation and experiment planning. No absolute benchmark numbers are published with the model card.
Deployment — tested with SGLang on 8 GPUs (--reasoning-parser glm45 --tool-call-parser glm47); keep generation_config.json since it lists the end-of-turn tokens.
License — MIT for the model weights; Apache-2.0 for independently authored Deep Principle code/documentation (see LICENSING.md/NOTICE). Contact: info@deepprinciple.com. Tech blog: https://blog.deepprinciple.com/introducing-hydrogen-1.0/
Training Data Post-training (SFT + RL) on domain data from chemistry, materials science, and life sciences, task-specific scientist feedback, laboratory records, and discovery trajectories; full datasets and training recipes not released.
Hardware Requirements
Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 43.0B are active per token
| Quantization | Weights | KV Cache | Total VRAM |
|---|---|---|---|
| FP16 — 16-bit | 1506.7 GB | 6.7 GB | 1588.7 GB |
| INT8 — 8-bit | 753.3 GB | 3.3 GB | 794.3 GB |
| INT4 — 4-bit | 376.7 GB | 1.7 GB | 397.2 GB |
| FP4 — 4-bit | 376.7 GB | 1.7 GB | 397.2 GB |
Architecture
- Attention
- Sparse Attention
- MoE
- yes
- Context
- 1M tokens
- Parameters
- 753329.9M
- Active params
- 43046.5M
Source: extracted model record · model repo
zai-org/GLM-5.3 (753B total / 43B active)
full-model MoE SFT + RL on scientific domain data
Training Pipeline
-
1
sft
Supervised fine-tuning on scientific domain data
Blog "How we train Hydrogen-1.0": pipeline combines supervised fine-tuning (SFT) and reinforcement learning (RL). SFT provides worked examples of reasoning, experiment planning, protocol design, and results analysis, drawing on domain knowledge (chemistry, materials science, life sciences) and full-cycle laboratory records. Full-model MoE training on Megatron + SGLang across CUDA and ROCm; offline packing retains complete reasoning examples without truncation. Hyperparameters not published.
-
2
rl
Reinforcement learning on research outcomes and discovery trajectories
Blog/README: RL trains the model to make better decisions using feedback on research outcomes — task-specific feedback from scientists, laboratory records, and complete discovery trajectories (e.g. materials discovery toward specific property targets like HOMO-LUMO gap). Resumable alternating rollout/training on shared GPUs with token-aligned behavior log-probability correction; BF16 and 128x128 blockwise FP8 inference weight export. Hyperparameters not published.
No comments yet
Be the first to share your thoughts!