Nahual puts text and 3D molecules on one sequence: a single set of weights handles 12 chemistry tasks and tops DiffCSP on crystal structure prediction
Synopsis
Nahual, released by a University of Toronto-led group, writes text and 3D molecules into one autoregressive sequence and uses next-token diffusion so each atom coordinate is denoised while attending only to the clean tokens before it. Trained jointly on 12 chemistry tasks into one set of weights, it reaches 79.9% MatchRate@20 on MP-20 crystal structure prediction against 77.9% for DiffCSP, and reports state-of-the-art QM9 property-conditional generation.
Interpretation
A decoder-only autoregressive model can accept and emit text and 3D molecules of any length, covering 12 chemistry task families from transition-state search to adsorption on metal surfaces with one set of weights. Previously, language models handled text questions while specialized models predicted 3D structures for molecules and crystals, with agentic systems calling the specialists as tools; text-only models serialized coordinates as digit strings, losing 3D continuity and inflating context length. On MP-20 crystal structure prediction, Nahual reaches 79.9% MatchRate@20 versus 77.9% for the specialized diffusion model DiffCSP; on QM9 property-conditional generation, self-reward sampling beats JODO and PropMolFlow on all six properties.
Each atom is represented as a discrete element type plus a continuous 3D coordinate, and the coordinate is produced by diffusion denoising rather than split into digit tokens. Earlier per-token diffusion heads saw earlier positions only through a fixed-size conditioning vector, while diffusion forcing trained the denoiser on noisy contexts. Next-token diffusion has a causal transformer denoise each coordinate while attending to clean preceding tokens, and inference reuses standard KV caching; against a per-token MLP diffusion head it is more accurate on transition states, noncovalent dimers, microsolvation, crystals, conformers and alloy surfaces, and on par on metal complexes and MOF adsorption.
Perspective
Evaluation covers 12 task families, including transition-state search, noncovalent interactions, lanthanide complexes, microsolvation, crystal structure prediction, conformer search, adsorption on metal surfaces and in porous materials, MOF and molecular property prediction, conditional molecule generation and NEB trajectories; most evaluations are in-distribution and reuse established splits. The authors propose next steps of adding more supervised and unsupervised geometry data, reinforcement learning on verifiable tasks, and few-step diffusion to cut inference cost.
The authors report that Nahual catastrophically forgets natural language, that each task uses a single fixed prompt template so the model learns task schemas rather than robust instruction following, and that capacity is unevenly spread: the largest tasks ODAC25 and GEOM-DRUGS are underfit while NCI and QCG-10 begin to overfit. Numerical property prediction lags because digits are supervised with cross-entropy; on ODAC25 the median adsRMSD grows from below 1 Å to over 4 Å above roughly 100 atoms, with steric clashes the most common failure; inference costs substantially more than specialized baselines because each atom requires a multi-step diffusion ODE through the second stack.
