重點導讀
Deep Principle 推出首款用於科學研究的開放權重模型 Hydrogen-1.0,並回報其在 FrontierScience-Research 的成績高於 GPT-6-Astra。
- 訓練方式:以 GLM-5.3 為基礎,使用自有科學資料、實驗室紀錄與發現歷程進行後訓練。
- 評測成績:Hydrogen-1.0 為 38.3%,GPT-6-Astra 為 33.3%。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
elvis@omarsar0
·20 天前
This is brilliant. Finally, we have a strong new open model for science.
As I said before, nothing beats a specialized model, and I think this will be even more important for science.
Hydrogen-1.0 is Deep Principle's first open-weight model for scientific research. They post-trained GLM-5.3 on their own scientific data, lab records, and discovery trajectories.
They report 38.3% on FrontierScience-Research, compared with 33.3% for GPT-6-Astra.
We need more of this.
推文來源 · X (Twitter)在 X 參與討論