Zatom-2 launches: one atomistic generative model spanning molecules, materials and proteins, fully open-source
rishabh16_ · x · 2026-10-11
Zatom-2, a collaboration across Cambridge, Yale, MIT and Lawrence Berkeley National Laboratory, is out: an atomistic multitask generative model covering molecular, material, and protein science in a single model. Paper on arXiv (2610.11454), with code, Hugging Face checkpoints, and interactive 3D visualizations all open.
- Method: multitask pretraining on atomistic data with flow matching—mass is transported along continuous velocity fields from Gaussian coordinate noise to stable chemical valence, with interactive denoising trajectories for molecules (OMol25), periodic materials (OMat24), and proteins (SCOPe).
- New capabilities: after finetuning it can generate all-atom proteins by SCOPe class; the release ships easter eggs honoring Zatom-1 plus the nucleus and zevals companion tools.
- Team: lead author Miruna Cretu (Cambridge), senior author Alex Morehead (LBNL/UC Berkeley).
A notable open AI-for-Science release in single-model cross-domain generative modeling.
More from Research
- 228 tasks, 26 models: METR's time horizon curve is nearly flat from 2-30 min — lulzxdxdxd · 2026-10-11
- mtDIVE preprint maps functional effects of thousands of mitochondrial DNA variants — manorlaboratory · 2026-10-11
- CSCW paper: only 20% of 1,800+ Mastodon instances share blocklists, few explain why — manoelribeiro · 2026-10-11
- 500B Tokens Later: Letting AI Agents Decompile a First-Person Shooter — davikr · 2026-10-11
- 'Alien math': AI proofs are correct but think in ways no human would — imjustnewatai · 2026-10-11
- AI will solve math, but Steve Strogatz says deciding which problems matter stays human — stevenstrogatz · 2026-10-11