Multi-Modal Foundation · mist-models
MIST 28M · QM8
Predicts electronic spectra properties for small organic molecules from SMILES strings using a 28M transformer.
What MIST 28M · QM8 is
MIST 28M QM8 is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It predicts electronic transition properties of small organic molecules directly from simplified molecular-input line-entry system (SMILES) chemical notations.
The model forms part of the Molecular Insight SMILES Transformers family. Pre-trained on synthetic organic compounds and fine-tuned on benchmark quantum chemical data, it provides computational chemists and materials scientists with rapid approximations of excitation energies and oscillator strengths without requiring resource-intensive quantum chemical simulations.
How it works
The underlying architecture is an 8-layer RoBERTa-PreLayerNorm encoder with 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a context window of 2048 tokens. The encoder was initially pre-trained on Kekulized SMILES strings from the Enamine REAL Space collection using a masked language modelling objective, cross-entropy loss, and the FusedLAMB optimizer.
For property prediction, the encoder is paired with a task network consisting of a two-layer multi-layer perceptron with Gaussian Error Linear Unit (GELU) activations and dropout. Hidden states are pooled using the representation of the first token. The full architecture was fine-tuned on the QM8 dataset via multi-task regression using mean squared error loss and the AdamW optimizer to predict 12 continuous properties.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with 2-layer MLP task head
- Parameter count
- 28 million
- Pre-training dataset
- Enamine REAL Space
- Fine-tuning dataset
- QM8 (MoleculeNet, 21,786 molecules)
- Context length
- 2,048 tokens
- Input format
- Kekulized SMILES strings
- Output format
- 12 scalar regression predictions or mean-pooled embedding vectors
What you can use it for
- Predicting Electronic Excitation Energies
- Calculates eight electronic excitation energies computed using time-dependent density functional theory for small organic molecules.
- Estimating Optical Oscillator Strengths
- Yields four oscillator strength values per molecule to assess transition probabilities and absorption intensity.
- Molecular Feature Extraction
- Extracts pooled numerical embedding vectors from molecular SMILES sequences for downstream clustering or property analysis.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
MIST 28M · QM8
Strengths and limits
Good at
- Predicts 12 electronic transition properties simultaneously within a single multi-task regression forward pass.
- Pre-trained on Enamine REAL Space, grounding representations in synthetically accessible chemical space.
- Lightweight 28M parameter footprint allows fast inference and low compute requirements.
Watch out for
- Requires Kekulized SMILES strings rather than generic or non-canonicalised molecular structures.
- Restricted to the chemical scope and property ranges present within the small-molecule QM8 benchmark.
- Relies on the SMIRK tokenization package, which requires a Rust environment to compile and execute.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use MIST 28M · QM8 commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run MIST 28M · QM8 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
Which molecular properties does the model predict?
The model outputs 12 scalar targets: eight electronic excitation energies and four oscillator strengths, as calculated with time-dependent density functional theory (TDDFT).
What input format does the model expect?
It expects Kekulized SMILES strings representing molecular structures, processed using the SMIRK tokenizer.
How was the model evaluated during development?
Performance was assessed using an 80/10/10 random split of the QM8 dataset, measuring the Mean Absolute Error (MAE) averaged across all 12 target properties.
Can the model produce latent representations instead of scalar values?
Yes, the encoder produces sequence embeddings that can be pooled to generate fixed-length molecular feature vectors.
How to cite it
Wadell, A., Bhutani, A., Azumah, V., Ellis-Mohr, A. R., Kelly, C., Zhao, H., Nayak, A. K., Hegazy, K., Brace, A., Lin, H., Emani, M., Vishwanath, V., Gering, K., Alkan, M., Gibbs, T., Wells, J., Varshney, L. R., Ramsundar, B., Duraisamy, K., Mahoney, M. W., Ramanathan, A., & Viswanathan, V. (2025). Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics]. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 28M · QM8's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.