Nahual: a decoder-only model unifying text and 3D molecules across 12 tasks
CatAstro_Piyush · x · 2026-11-17
Nahual is a decoder-only model that processes interleaved sequences of text and 3D molecules directly, removing the need to have LLMs call external tools for molecular structure. Atoms are serialized into <xyz>-delimited coordinate tokens and trained jointly with text.
Per the project figure, one unified model handles 12 tasks: transition state search, noncovalent interactions, metal complexes, microsolvation, crystal structure prediction, molecular conformer search, adsorption on metal surfaces and in porous materials, MOF and molecular property prediction, conditional molecule generation (e.g. 6846.4 meV), and nudged elastic band trajectories.
Accepted at NeurIPS 2026, the work points toward unifying molecular simulation tasks within an LLM-style sequence modeling framework.
More from Research
- MIT-led WorldParticle: one AI model simulates sand, fluids, and protein folding — MacrinePhD · 2026-11-17
- Sage releases open S2.1 Flash science multimodal model, tops MLE-bench Lite at 90.3 — QuixiAI · 2026-11-17
- Anthropic details four categories of unintended Claude actions, from paywall bypass to URL tricks — JustinLin610 · 2026-11-17
- Edge ML engineer questions why vanilla ViTs beat hybrid CNN+Attention stacks — terrrp · 2026-11-17
- NVIDIA open-sources cuPotential, over 3x faster MD nonbonded interactions vs OpenMM — AllThingsApx · 2026-11-17
- Replicating 4-bit forward-pass training on one B200: 38% faster at 3.3B with identical loss — OtherRaisin3426 · 2026-11-17