Deep Principle发布Hydrogen-1.0开源科学模型
Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery
专注垂直领域的开源大模型发布,提供完整权重与详细对比数据,适合科研场景复现与评测,值得开发者关注。
Deep Principle has released Hydrogen-1.0, an open-weight scientific model for chemistry, materials science and life sciences. It is built by post-training Z.ai’s GLM-5.3 on domain data, laboratory records and agent discovery trajectories. For developers, it offers open weights that report a FrontierScience-Research win over GPT-6-Astra. The weights ship under the MIT License.
Deep Principle 发布了 Hydrogen-1.0,这是一个面向化学、材料科学和生命科学的开源权重科学模型。它是在领域数据、实验室记录和智能体发现轨迹上对 Z.ai 的 GLM-5.3 进行后训练构建而成。对于开发者而言,它提供了开源权重,据称在 FrontierScience-Research 基准测试中击败了 GPT-6-Astra。该权重采用 MIT 许可证发布。
TL;DR
TL;DR(太长不看)
- Size: 753B parameters (MoE, ~40B active per the GLM-5.3 base), 1M-token context, 128K max output.
- Runs on: 8 GPUs with SGLang (tested); NVFP4 for NVIDIA B200/B300, MXFP4 for AMD MI350X/MI355X.
- Performance: Beats GPT-6-Astra on FrontierScience-Research; roughly matches Kimi-K3 on GPQA-Diamond.
- Best: 38.3% on FrontierScience-Research vs 33.3% for GPT-6-Astra (self-reported).
- Worst: Trails GPT-6-Astra (max thinking) by 3.33 points on NMR structure elucidation.
- 规模:753B 参数(MoE 架构,根据 GLM-5.3 基座,每次激活约 40B),1M-token 上下文,最大输出 128K。
- 运行环境:使用 SGLang 在 8 张 GPU 上运行(已测试);支持 NVIDIA B200/B300 的 NVFP4 格式,以及 AMD MI350X/MI355X 的 MXFP4 格式。
- 性能表现:在 FrontierScience-Research 基准测试中击败 GPT-6-Astra;在 GPQA-Diamond 基准测试中与 Kimi-K3 大致持平。
- 最佳表现:在 FrontierScience-Research 基准测试中获得 38.3% 的成绩,而 GPT-6-Astra 为 33.3%(自我报告)。
- 最差表现:在 NMR 结构解析任务中,落后于 GPT-6-Astra(最大思考模式)3.33 分。
What is Hydrogen-1.0?
什么是 Hydrogen-1.0?
Hydrogen-1.0 is a scientific research assistant model from Deep Principle, a Hangzhou-based AI-for-materials company. The model card describes it as a GLM-5.3 derivative for scientific problem solving. It targets question framing, experiment planning, results analysis and open-ended discovery. Deep Principle team states that proprietary models produced recent AI science results, but their discovery capabilities are not open. Hydrogen-1.0 is its attempt at a reproducible, inspectable alternative.
Hydrogen-1.0 是 Deep Principle 推出的一款科学研究辅助模型。Deep Principle 是一家位于杭州的 AI for Materials 公司。模型卡片将其描述为用于解决科学问题的 GLM-5.3 衍生版本。它旨在针对问题构建、实验规划、结果分析和开放式发现等任务。Deep Principle 团队指出,其专有模型近期取得了 AI 科学领域的成果,但其发现能力并未开放。Hydrogen-1.0 是其尝试提供的一种可复现、可检查的替代方案。
How was Hydrogen-1.0 trained?
Hydrogen-1.0 是如何训练的?
The pipeline combines supervised fine-tuning (SFT) with reinforcement learning (RL). Data comes from 4 sources:
训练流程结合了监督微调(SFT)与强化学习(RL)。数据来自以下 4 个来源:
- Domain knowledge: chemistry, materials science and life sciences data.
- Research scenarios: task-specific feedback from Deep Principle scientists.
- Laboratory records: full-cycle electronic lab notebook records captured through Mira.
- Discovery trajectories: sequences of choices and feedback, including failed attempts.
- 领域知识:化学、材料科学和生命科学数据。
- 研究场景:来自 Deep Principle 科学家的特定任务反馈。
- 实验室记录:通过 Mira 捕获的全周期电子实验笔记本记录。
- 发现轨迹:包括失败尝试在内的选择序列和反馈序列。
Infrastructure builds on Megatron and SGLang across NVIDIA CUDA and AMD ROCm. The team keeps dynamic sparse attention causal across context-parallel shards. Muon optimizer state is moved between CPU, GPU and disk per backend. RL uses resumable rollouts, and each token keeps its original behavior log-probability for importance correction. A layout adapter converts training checkpoints into BF16 and 128×128 blockwise FP8 inference weights. Full datasets and recipes are not released.
基础设施基于 Megatron 和 SGLang,跨 NVIDIA CUDA 和 AMD ROCm 平台构建。团队保持动态稀疏注意力机制在上下文并行分片上的因果一致性。Muon 优化器状态根据后端情况在 CPU、GPU 和磁盘之间移动。强化学习使用可恢复的 rollout,每个 token 保留其原始行为的 log-probability 以进行重要性校正。布局适配器将训练检查点转换为 BF16 和 128×128 块状 FP8 推理权重。完整数据集和配方未发布。
How does Hydrogen-1.0 perform on benchmarks?
Hydrogen-1.0 在基准测试中的表现如何?
On public benchmarks, Deep Principle reports Hydrogen-1.0 outperforms GPT-6-Astra on FrontierScience-Research. The tweet thread from one of the founders’ report 93.0% on GPQA-Diamond. That improves on GLM-5.3 and roughly matches Kimi-K3, per the blog.
在公开基准测试中,Deep Principle 报告 Hydrogen-1.0 在 FrontierScience-Research 上优于 GPT-6-Astra。其中一位创始人的推文线程报告其在 GPQA-Diamond 上得分 93.0%。据博客所述,这优于 GLM-5.3,并大致与 Kimi-K3 持平。
Internal research-task evals compare it with GPT-6-Astra at maximum thinking:
内部研究任务评估将其与最大思考模式下的 GPT-6-Astra 进行比较:
- Scientific computing: Hydrogen-1.0 leads by 5.13 percentage points.
- Forward reaction prediction: Hydrogen-1.0 leads by 4.76 percentage points.
- NMR structure elucidation: GPT-6-Astra leads by 3.33 points.
- 科学计算:Hydrogen-1.0 领先 5.13 个百分点。
- 正向反应预测:Hydrogen-1.0 领先 4.76 个百分点。
- 核磁共振结构解析:GPT-6-Astra 领先 3.33 分。
In transition metal complex design, models chased a larger HOMO-LUMO gap over 5 rounds. By round 5, Hydrogen-1.0 reached a mean best-so-far gap of 4.318 eV. GPT-6-Astra reached 3.689 eV and Kimi-K3 reached 3.353 eV. On 117 experiment-planning questions, Hydrogen-1.0 scored above Kimi-K3. GPT-6-Astra and Opus-5.5 were excluded after refusing more than 50% of requests.
在过渡金属配合物设计中,模型在多轮迭代中追求更大的 HOMO-LUMO 能隙。到第 5 轮时,Hydrogen-1.0 达到平均最佳迄今能隙为 4.318 eV。GPT-6-Astra 达到 3.689 eV,Kimi-K3 达到 3.353 eV。在 117 个实验规划问题上,Hydrogen-1.0 的得分高于 Kimi-K3。由于拒绝超过 50% 的请求,GPT-6-Astra 和 Opus-5.5 被排除在外。
What can it do in a real lab workflow?
它在真实实验室工作流中能做什么?
The technical post shows 4 demos. Hydrogen-1.0 writes Python to build a polyethylene melt and runs GROMACS. It detects a topology error, repairs it, and analyzes the corrected 40 ns trajectory. It grows an Ireland-Claisen reaction network of 27 structures through 30 plausible reactions. It ranks 7 candidate monoformylation products of 2-bromonaphthalene from a 1H NMR trace. It also writes a JSON protocol for a 12-vial Suzuki coupling study. A virtual lab compilation step catches an aspiration without a fresh tip before execution.
技术文章展示了 4 个演示案例。Hydrogen-1.0 编写 Python 代码构建聚乙烯熔体并运行 GROMACS。它检测到一个拓扑错误,修复了该错误,并分析了校正后的 40 ns 轨迹。它通过 30 种合理的反应生成了一个包含 27 种结构的 Ireland-Claisen 反应网络。它根据 1H NMR 谱图对 2-溴萘的 7 种候选单甲酰化产物进行排名。它还编写了一个用于 12 瓶 Suzuki 偶联研究的 JSON 协议。一个虚拟实验室编译步骤在执行前捕获了一个缺少新鲜吸头的情况。
How do you deploy Hydrogen-1.0?
如何部署 Hydrogen-1.0?
Weights are on Hugging Face and ModelScope. Deep Principle team tested BF16 serving with SGLang on 8 GPUs using tensor parallelism of 8. NVFP4 targets NVIDIA B200 and B300. MXFP4 targets AMD Instinct MI350X and MI355X. The server exposes an OpenAI-compatible API with a reasoning_effort setting.
权重文件托管在 Hugging Face 和 ModelScope 上。Deep Principle 团队使用 SGLang 在 8 块 GPU 上以张量并行度为 8 的方式测试了 BF16 服务。NVFP4 针对 NVIDIA B200 和 B300。MXFP4 针对 AMD Instinct MI350X 和 MI355X。服务器提供与 OpenAI 兼容的 API,并支持 reasoning_effort 设置。
Hydrogen-1.0 vs GLM-5.3, Kimi K3 and GPT-6-Astra
Hydrogen-1.0 与 GLM-5.3、Kimi K3 和 GPT-6-Astra 的对比
| Feature | Hydrogen-1.0 | GLM-5.3 | Kimi K3 | GPT-6-Astra |
|---|---|---|---|---|
| Developer | Deep Principle | Z.ai | Moonshot AI | OpenAI |
| Total params | 753B | 753B (HF) | 2.8T | Not disclosed |
| Active params | ~40B (inherited) | 40B (LM Studio) | 16 of 896 experts | Not disclosed |
| Context | 1M | 1M | 1M | 1.05M |
| Weights / license | Open, MIT | Open, GLM-5.3 License | Open | Proprietary |
| Modality | Text | Text | Text + vision | Not disclosed here |
| Science focus | Chemistry, materials, life sciences | Coding, agents | General | General |
| FrontierScience-Research | 38.3% | Not disclosed | Not disclosed | 33.3% (per Deep Principle) |
| GPQA-Diamond | 93.0% | Not disclosed | Not disclosed | 96.0% (OpenAI-reported) |
| API price (in/out per 1M) | NeuroSpark: Not disclosed | $1.40 / $4.40 (Z.ai) | $3.00 / $15.00 | $10.00 / $50.00 (OpenAI) |
| 特性 | Hydrogen-1.0 | GLM-5.3 | Kimi K3 | GPT-6-Astra |
|---|---|---|---|---|
| 开发者 | Deep Principle | Z.ai | Moonshot AI | OpenAI |
| 总参数量 | 753B | 753B (HF) | 2.8T | 未公开 |
| 活跃参数量 | ~40B(继承) | 40B(LM Studio) | 896个专家中的16个 | 未公开 |
| 上下文窗口 | 1M | 1M | 1M | 1.05M |
| 权重/许可证 | 开源,MIT | 开源,GLM-5.3 许可证 | 开源 | 专有 |
| 模态 | 文本 | 文本 | 文本 + 视觉 | 此处未公开 |
| 科学侧重 | 化学、材料、生命科学 | 编码、智能体 | 通用 | 通用 |
| FrontierScience-Research | 38.3% | 未公开 | 未公开 | 33.3%(据 Deep Principle 报告) |
| GPQA-Diamond | 93.0% | 未公开 | 未公开 | 96.0%(OpenAI 报告) |
| API 价格(每百万输入/输出) | NeuroSpark:未公开 | $1.40 / $4.40(Z.ai) | $3.00 / $15.00 | $10.00 / $50.00(OpenAI) |
Key Takeaways
关键要点
- Hydrogen-1.0 is an MIT-licensed, 753B GLM-5.3 post-train for science.
- It reports 38.3% on FrontierScience-Research, above GPT-6-Astra’s 33.3%.
- Training adds lab records and discovery trajectories, not just papers.
- It still trails GPT-6-Astra on NMR structure elucidation.
- NVFP4 and MXFP4 builds target Blackwell and MI350-class GPUs.
- Hydrogen-1.0 是一款采用 MIT 许可证的、面向科学领域的 753B GLM-5.3 后训练模型。
- 它在 FrontierScience-Research 上取得了 38.3% 的成绩,高于 GPT-6-Astra 的 33.3%。
- 其训练不仅包含论文,还加入了实验室记录和发现轨迹。
- 在核磁共振结构解析方面,它仍落后于 GPT-6-Astra。
- NVFP4 和 MXFP4 构建版本针对 Blackwell 和 MI350 级 GPU 进行优化。
Check out the Technical blog, Model weights on Hugging Face and Usage notes. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
请参阅技术博客、Hugging Face 上的模型权重以及使用说明。此外,欢迎在 Twitter 上关注我们,别忘了加入拥有超过 15 万成员的 ML SubReddit,并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力