Authors: Austin H. Cheng, Marcel Müller, Andreas Burger, Yeonghun Kang, Tsz Wai Ko, Marta Skreta, Luka Mucko, Ella Miray Rajaonson, Cher Tian Ser, Jérôme F. Gonthier, Alex Zook, Varinia Bernales, Alán Aspuru-Guzik
Organizations: University of Toronto, Toronto, ON, Canada · Vector Institute for Artificial Intelligence, Toronto, ON, Canada · NVIDIA · Sungkyunkwan University, Republic of Korea · Université de Montréal, Montréal, QC, Canada · Mila – Québec AI Institute, Montréal, QC, Canada · University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, Croatia · Canadian Institute for Advanced Research (CIFAR), Toronto, ON, Canada · Acceleration Consortium, Toronto, ON, Canada
Generative models play a rapidly growing role in the design and simulation of molecules. The most widely used applications of generative models are in large language models (LLMs), owing to the natural flexibility of language for specifying and answering queries. However, generalist LLMs do not effectively handle large sets of 3D coordinates, whereas specialist generative models of 3D molecules cannot reach the controllability afforded by language. To bridge these capabilities, we propose Nahual, a decoder-only autoregressive diffusion model for any sequence of text and 3D molecules. To enable expressive autoregression on continuous coordinates, we propose next-token diffusion, whereby a causal transformer denoises each token while attending to the clean tokens before it, without increasing context size. We train Nahual on a curated set of twelve 3D chemistry tasks ranging from microsolvation to molecular conformer search to adsorption on metal surfaces, with associated evaluation metrics and baselines. The result is a single set of end-to-end-trained weights that simultaneously performs all tasks and surpasses state-of-the-art models specialized for crystal structure prediction and property-conditional molecule generation. Our results demonstrate the viability of scaling generalist decoder-only models for native multimodal understanding and generation across language and chemistry. Code is available at https://github.com/the-matter-lab/nahual.
Figures & tables
Figure 1: Nahual processes any sequence of text and 3D molecules via next-token prediction. A token (a,x) has a discrete part a (a word ID or atomic number) and a continuous part x∈R3 (zero-padding for words, atomic position for atoms). Nahual ’s general structure allows it to be trained simultaneously on 12 different 3D chemistry tasks.
Figure 2: A single instance of Nahual solves a diverse set of 3D chemistry tasks: transition-state search, noncovalent dimer assembly, building lanthanide coordination complexes, microsolvation with a variable number of waters, inorganic crystal structure prediction, and molecular conformer search. Prompts are shown verbatim.
Task
Dataset
Metric
Baseline method
Baseline
Nahual
Structure search and assembly
Transition state search
Transition1x
median RMSD@1 ↓
TSDiff
0.221
0.197 ± 0.013
Noncovalent interactions
NCI Atlas
median RMSD@10 ↓
CREST
0.368
0.335 ± 0.039
Metal complexes
LnQM
ConnectivityMatch@1 ↑
Architector
10.2
73.3 ± 1.1
Microsolvation
QCG-10
EwT@10 ↑
Bulk-water cutout
18.6
50.6 ± 0.2
Crystal structure prediction
MP-20
MatchRate@20 ↑
DiffCSP
77.9
79.9 ± 0.1
Table 1: A single generalist Nahual versus per-task baselines. Best per row in bold . Lower is better for ↓ and higher for ↑ . Root-mean-square deviation (RMSD) is in Å. For “@ k ”, we sample k responses per test example and take the best score. Nahual entries report mean ± standard error over two training seeds. Task details, metric definitions, and more baselines are in Appendix G .
Figure 3: A single instance of Nahual solves a diverse set of 3D chemistry tasks: placing adsorbates on metal surfaces and in porous MOFs, MOF and molecular property prediction, conditional molecule generation, and nudged elastic band initialization. Prompts are shown verbatim.
Figure 4: (a) QM9 property-conditional generation, scored by EGNN-MAE (lower is better), for GPT-5.4, the EGNN variant of PropMolFlow ( Table 9 ), single-sample Nahual , and Nahual with self-reward@5. (b) COV-R on GEOM-DRUGS versus RMSD threshold, mean (blue) and median (orange) over the 932 test molecules, for Nahual (solid), DiTMC+PE(3)-L Frank et al. [2025] (dashed), and DMT-L Liu et al. [2025] at 0.75 Å (stars). (c) QCG-10 EwT@10 by number of waters N . Nahual results are means over two seeds.
Figure 5: Next-token diffusion versus DiffMLP over training , on fixed 128 -prompt test subsets, so values differ from Table 1 . Lines are means over two seeds, with shading showing the span. RMSDs are medians and QM9 cond. uses self-reward@5. Figure 7 shows all tasks.
Appendix figures & tables37 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Next-token diffusion. Diffusion Loss Li et al. [2024] equips a transformer with a lightweight per-token diffusion head, but conditioning is bottlenecked through a fixed-size vector. Diffusion Forcing Chen et al. [2024] lets the denoiser attend to the full sequence, but in training it sees only noisy context. Our proposed next-token diffusion grounds the transformer in the clean previous sequence and denoises only the next sequence.
Figure 7: Evaluation over training for next-token diffusion and DiffMLP ( Section B.1 ), two seeds each, every 5,000 steps from 5 K to 115 K. Each task is evaluated on a fixed subset of at most 128 test prompts ( 10 molecules for GEOM-DRUGS), so values differ from Table 1 . Crosses mark non-finite metrics from failed generations. OC20NEB pools the three reaction types.
Task
Dataset
Metric
Next-token diffusion
DiffMLP
Structure search and assembly
Transition state search
Transition1x
median RMSD@1 ↓
0.197 ± 0.013
0.247 ± 0.004
Noncovalent interactions
NCI Atlas
median RMSD@10 ↓
0.335 ± 0.039
0.442 ± 0.009
Metal complexes
LnQM
ConnectivityMatch@1 ↑
73.3 ± 1.1
73.5 ± 0.3
Microsolvation
QCG-10
EwT@10 ↑
50.6 ± 0.2
43.0 ± 1.4
Crystal structure prediction
MP-20
MatchRate@20 ↑
79.9 ± 0.1
76.5 ± 0.3
Appendix
Table 2: Next-token diffusion versus a per-token MLP denoiser (DiffMLP). The models differ only in the denoiser. Entries are mean ± standard error over two training seeds under the protocol of Table 1 . The better mean is in bold . † Includes one malformed DiffMLP prediction of 822 eV. Without it, the DiffMLP MAE is 0.467 eV.
Dataset
# examples
weight wi
probability
MP-20 (raster order)
27,131
0.30
0.046
MP-20 (default order)
27,131
0.30
0.046
NCI Atlas
2,667
1.50
0.023
Transition1x
8,999
1.10
0.056
Transition1x (reactant/product swapped)
8,999
1.10
0.056
LnQM
8,784
1.00
0.050
Appendix
Table 3: Training data mixture used for the reported Nahual checkpoint.
Figure 8: Sequence-length distribution of the training mixture in tokens. The training data contain N=1,396,153 sequences. Blue: unweighted raw corpus. Red: weighted, the distribution seen during training. Top: linear scale. Bottom: log scale.
Task / dataset
Size
T1x
1,073
NCI Atlas
149
LnQM
488
MP-20
9,046
QCG-10
500
Alloy
1,126
Appendix
Table 4: Test-set sizes used for evaluation.
Task / dataset
Wall-clock
T1x
3.1 min
NCI Atlas
1.1 min
LnQM
3.0 min
MP-20
64 min
QCG-10
16 min
Alloy
3.7 min
Appendix
Table 5: Inference cost of each evaluation for one checkpoint, as node-equivalent wall-clock on a 4 × GB200 node (total GPU time divided by four).
Figure 9: T1x transition-state RMSD on the test set of Duan et al., as per-prompt violins (top) and CDF (bottom) on a shared log-RMSD axis.
Figure 10: T1x transition-state predictions. Nahual prediction (cyan) overlaid on the reference.
Figure 11: NCI Atlas dimer RMSD: per-method box (top) and per-prompt CDF (bottom) on a shared log-RMSD axis.
Figure 12: NCI Atlas dimer RMSD: best-of- k scaling vs. sample budget. Left: mean. Right: median.
Figure 13: NCI Atlas dimer assembly on three test prompts (one per column). Top row: Nahual prediction (cyan) overlaid on the reference dimer. Bottom row: CREST prediction (yellow) overlaid on the same reference dimer.
Figure 14: LnQM connectivity match rate: Nahual ( k=1 ) vs. Architector pipeline stages.
Figure 15: LnQM lanthanide complexes built from the natural-language description. Nahual prediction (cyan) overlaid on the LnQM ground-truth geometry.
Figure 16: Energy within Threshold (EwT) at k=1 , k=3 , and k=10 by number of waters N . Nahual with xTB relaxation beats the bulk-water cutout baseline at every N . Nahual bars are means over two seeds, and open circles mark bars of height zero.
Figure 17: Microsolvation (QCG-10) on three test prompts (one per column). Top row: xTB-CREST reference cluster. Bottom row: the Nahual prediction.
Figure 18: MP-20 crystal structure prediction: best-of- k MatchRate scaling for Nahual against DiffCSP’s published k=1 / k=20 numbers.
Figure 19: MP-20 crystal structure prediction. Nahual ’s predicted unit cells, aligned to the Materials Project reference with StructureMatcher .
Method
COV-R [%] ↑
AMR-R [Å] ↓
COV-P [%] ↑
AMR-P [Å] ↓
Mean
Median
Mean
Median
Mean
Median
Mean
Median
GeoMol (0.3M)
44.6
41.4
0.875
0.834
43.0
36.4
0.928
0.841
GeoDiff (1.6M)
42.1
37.8
0.835
0.809
24.9
14.5
1.136
1.090
Tors. Diff. (1.6M)
72.7
80.0
0.582
0.565
55.2
56.9
0.778
0.729
MCF-L (242M)
84.7
92.2
0.390
0.247
66.8
71.3
0.618
0.530
DMT-L (150M)
85.8
92.3
0.375
0.346
67.9
72.5
0.598
0.527
Appendix
Table 6: GEOM-DRUGS results, with parameter counts in parentheses. -R is recall and -P is precision. Best in bold , second best underlined . Baseline numbers are from the original papers. All methods use the 1,000 -molecule test split of GeoMol and Torsional Diffusion. We follow the evaluation protocol of Torsional Diffusion, which skips molecules whose reference conformers do not match their SMILES, keeping only 932 molecules. Nahual is the mean over two seeds.
Figure 20: GEOM-DRUGS coverage versus threshold δ for Nahual on the 932 scored test molecules. Left: COV-R, the fraction of reference conformers within δ Å of some generated conformer. Right: COV-P, the fraction of generated conformers within δ of some reference. Blue is the mean over molecules and orange is the median. Nahual (solid) is averaged over two seeds. DiTMC+PE(3)-L (dashed) is extracted from the plots of Frank et al. [2025] , which stop at 1.5 Å. Stars mark the published performance of DMT-L at the standard threshold of 0.75 Å (dotted line).
Figure 21: Median best-of- k adsorbate RMSD vs. sample budget per subset (Alloy / FG / OCD).
Figure 22: Median best-of- k full-system RMSD vs. sample budget per subset (Alloy / FG / OCD).
Figure 23: Alloy on three test prompts at the 25/50/75th percentile of adsorbate RMSD (one per column). Top row: Nahual prediction at k=1 . Bottom row: ground-truth structure. Red/green/blue arrows mark the three lattice vectors and the wireframe box outlines the periodic cell.
Figure 24: FG on three test prompts at the 25/50/75th percentile of adsorbate RMSD (one per column). Top row: Nahual prediction at k=1 . Bottom row: ground-truth functional-group adsorbate. Lattice arrows and wireframe box mark the periodic cell.
Figure 25: OCD (OC20-Dense) on three test prompts at the 25/50/75th percentile of adsorbate RMSD (one per column). Top row: Nahual prediction at k=1 . Bottom row: ground-truth structure. Lattice arrows and wireframe box mark the periodic cell.
k=1
k=10
Diagnostic
Nahual
UMA
Nahual
UMA
Steric clash
15.0 ± 5.0%
1.1%
8.3 ± 2.4%
0.7%
Adsorbate exploded
3.1 ± 2.8%
4.1%
0.6 ± 0.2%
3.0%
Adsorbate floats far
3.3 ± 1.9%
5.6%
1.9 ± 0.0%
4.1%
Appendix
Table 7: Per-prompt failure-mode rates on the 270-prompt ODAC25 test split. Nahual entries are mean ± standard error over two training seeds. k=1 uses the first sample, and k=10 the sample with the lowest adsorbate RMSD. Steric clash : some adsorbate–framework distance is below 0.85× the sum of covalent radii. Adsorbate exploded : the largest intra-adsorbate distance exceeds the DFT reference maximum by more than 1 Å. Adsorbate floats far : the closest adsorbate–framework distance exceeds 4 Å.
Figure 33
Figure 28: ODAC25 adsorbate-in-MOF placements on three test prompts (one per column). Top row: Nahual prediction at k=1 . Bottom row: ground-truth structure. Each prompt has several reference geometries, and we show the first. Lattice arrows and wireframe box mark the periodic cell.
Figure 29: MOF property prediction parity plots. Rows: GPT-5.4-high (3-shot) and Nahual . Columns: QMOF bandgap, CoRE MOF DB surface area, and CoRE MOF DB pore volume. The dashed line is y=x .
Figure 30: QM9 property prediction MAE@1 (lower is better) for Nahual , GPT-5.4-high (3-shot), and the EGNN regressor.
Figure 31: QM9 property prediction parity plots for Nahual , one panel per property. The dashed line is y=x . The few predictions outside the axes are drawn at the border.
Property
α
Δϵ
ϵHOMO
ϵLUMO
μ
Cv
Units
Bohr 3
meV
meV
meV
Debye
cal/(mol ⋅ K)
EGNN regressor Satorras et al. [2021] , Bao et al. [2023]
0.0992
64.9
38.8
37.0
0.0452
0.0399
GPT-5.4-high (3-shot)
4.473
836
384
992
0.611
2.129
Nahual
0.523
164.5
125.9
135.8
0.459
0.245
Appendix
Table 8: QM9 property prediction MAE@1 on the 13,083-molecule test split. EGNN is the pretrained property predictor of EEGSDE. GPT-5.4-high (3-shot) numbers are those of Table 1 . Best per column in bold , lower is better.
Figure 32: QM9 conditional generation EGNN-MAE (lower is better). Nahual bars are means over two seeds, and PropMolFlow is its EGNN variant ( Table 9 ).
Property
α
Δϵ
ϵHOMO
ϵLUMO
μ
Cv
Units
Bohr 3
meV
meV
meV
Debye
cal/(mol ⋅ K)
QM9 (lower-bound)
0.10
64
39
36
0.043
0.040
Random (upper-bound)
9.01
1470
645
1457
1.616
6.857
# Atoms
3.86
866
426
813
1.053
1.971
GPT-5.4-high (3-shot)
11.35
612
328
1290
0.450
6.511
EEGSDE Bao et al. [2023]
2.62
542
302
496
0.858
1.037
Appendix
Table 9: QM9 conditional generation EGNN-MAE. Baselines from EEGSDE to PropMolFlow are taken from Zeng et al. [2026] . QM9 (lower-bound) is the EGNN’s error on ground-truth structures. Random and # Atoms are null baselines. Nahual ( k=1 ) scores the first sample. Nahual (self-reward@5) draws 5 samples per prompt, predicts the property of each itself, and keeps the one closest to the target. Best per column among generation methods in bold .
Figure 33: QM9 conditional generation, one test prompt for each of α , Δϵ , ϵHOMO , ϵLUMO , μ , and Cv . For each property, the Nahual -generated molecule ( Nahual ) is shown above the QM9 molecule with the same target value ( Ground truth ).
Figure 34: QM9 polarizability ( α ) sweep: Nahual -generated molecules for six target values of α , increasing from left to right. Molecules with larger α are typically larger, and the model recovers this trend without explicit supervision.
Figure 35: OC20NEB median full-system RMSD for all reactions combined and for each reaction type. Methods are linear interpolation (grey), MIC linear interpolation (lavender), and Nahual best-of-1 (blue) and best-of-3 (green).
Department of Pharmaceutical Sciences, University of Vienna · Christian Doppler Laboratory for Molecular Informatics in the Biosciences · Vienna Doctoral School of Pharmaceutical, Nutritional and Sport Sciences Josef-Holaubek-Platz 2, 1090 Vienna, Austria