HyFF Andante reports 29x UMA-S-1.2 MD speed: what would you test before switching?
I came across Tencent's HyFF release. The claim that got my attention was the 29x MD speedup, but I'd want to count the evaluation work before calling it a saving. I haven't run it.
The maintainers' README reports Andante at 29x UMA-S-1.2's MD speed, with force MAE of 7.65 versus 7.55 meV/Å. The speed test is a 2,586-atom solvated chorismate system on one NVIDIA H20, using 1 fs NVE with bond constraints and including neighbour-list rebuilds. Accuracy is measured separately on the full OMol25 validation set.
That separation caught my attention. I wouldn't treat a validation-set error and one system's throughput as enough to estimate the cost of moving an existing project. The README also says other GPUs are untested, which matters before assuming the speedup will carry over to the hardware someone already owns.
My instinct would be to start with a small system whose behaviour is already understood, keep the comparison settings explicit, and count the time spent checking the result as part of the migration cost.
For people who actually use ML potentials: what would be your first acceptance test here? Long-trajectory stability, behaviour outside the training distribution, performance on your own GPU, or something else? I'm especially interested in what would make you stop evaluating a model early, before spending days integrating it.
5 points · 2 comments
HyInv 11月2日开放了代码和权重：给定蛋白质骨架，生成氨基酸序列。196万参数，CPU也能跑。
官方注明，输入只收 PDB，手里的 mmCIF 得先转换。想复现结果，还要固定批大小；同一个随机种子，换批大小也会换结果。
Tencent-Hunyuan/Hyra-results
2 likes · 158 views
科研开始造工具。
@NFT_Chen：腾讯混元科研 Agent Hyra 2.0——三层 Agent 可跑数天至数周；示例含蛋白质模型加速、香山 RISC-V 优化、机器人旋梯手仿真；成果开源。页面：hy.tencent.com · 成果：Tencent-Hunyuan/hyra-results
判断：做 AutoResearch / 需要可复用科研工具链的人先看开源结果；只做日常聊天的可跳过。先复现公开实验，别直接接生产数据。
2 likes · 162 views
Tencent Hunyuan's Hyra research agent shipped new open science models 🔬 HyFF molecular force fields and two protein models, HyInv and HyPLM-Flash, are out with weights and code.
#AI #AIAgents #OpenSource
Tencent-Hunyuan/Hyra-results
23 views
Tencent Hunyuan released Hyra 2.0, a research agent built to make reusable scientific tools 🔬 It has already trained and open-sourced two protein models, HyInv and HyPLM-Flash, with weights and code.
#AI #AIAgents #Tencent
Tencent-Hunyuan/hyra-results
Learn more: hy.tencent.ai
More research results are being published on GitHub: Tencent-Hunyuan/hyra-results
2 likes · 56 views
Tencent-Hunyuan/Hyra-results 最怕的就是自己造尺子自己量😅
4 likes · 68 views
Mentions
HyFF Andante reports 29x UMA-S-1.2 MD speed: what would you test before switching? I came across Tencent's HyFF release. The claim that got my attention was the 29x MD speedup, but I'd want to count the evaluation work before calling it a saving. I haven't run it. The maintainers' README reports Andante at 29x UMA-S-1.2's MD speed, with force MAE of 7.65 versus 7.55 meV/Å. The speed test is a 2,586-atom solvated chorismate system on one NVIDIA H20, using 1 fs NVE with bond constraints and including neighbour-list rebuilds. Accuracy is measured separately on the full OMol25 validation set. That separation caught my attention. I wouldn't treat a validation-set error and one system's throughput as enough to estimate the cost of moving an existing project. The README also says other GPUs are untested, which matters before assuming the speedup will carry over to the hardware someone already owns. My instinct would be to start with a small system whose behaviour is already understood, keep the comparison settings explicit, and count the time spent checking the result as part of the migration cost. For people who actually use ML potentials: what would be your first acceptance test here? Long-trajectory stability, behaviour outside the training distribution, performance on your own GPU, or something else? I'm especially interested in what would make you stop evaluating a model early, before spending days integrating it.
HyInv 11月2日开放了代码和权重：给定蛋白质骨架，生成氨基酸序列。196万参数，CPU也能跑。 官方注明，输入只收 PDB，手里的 mmCIF 得先转换。想复现结果，还要固定批大小；同一个随机种子，换批大小也会换结果。 Tencent-Hunyuan/Hyra-results
科研开始造工具。 @NFT_Chen：腾讯混元科研 Agent Hyra 2.0——三层 Agent 可跑数天至数周；示例含蛋白质模型加速、香山 RISC-V 优化、机器人旋梯手仿真；成果开源。页面：hy.tencent.com · 成果：Tencent-Hunyuan/hyra-results 判断：做 AutoResearch / 需要可复用科研工具链的人先看开源结果；只做日常聊天的可跳过。先复现公开实验，别直接接生产数据。
Tencent Hunyuan's Hyra research agent shipped new open science models 🔬 HyFF molecular force fields and two protein models, HyInv and HyPLM-Flash, are out with weights and code. #AI #AIAgents #OpenSource Tencent-Hunyuan/Hyra-results
Tencent Hunyuan released Hyra 2.0, a research agent built to make reusable scientific tools 🔬 It has already trained and open-sourced two protein models, HyInv and HyPLM-Flash, with weights and code. #AI #AIAgents #Tencent Tencent-Hunyuan/hyra-results
Learn more: hy.tencent.ai More research results are being published on GitHub: Tencent-Hunyuan/hyra-results
Tencent-Hunyuan/Hyra-results 最怕的就是自己造尺子自己量😅