Multi-Modal Foundation · mist-models
MIST 1.8B · QM8
A 1.8-billion-parameter molecular foundation model fine-tuned to predict electronic spectra properties from SMILES strings.
What MIST 1.8B · QM8 is
MIST 1.8B · QM8 is a fine-tuned variant of the Molecular Insight SMILES Transformer (MIST) family, created by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It is designed to perform regression predictions for electronic transitions in organic molecules directly from simplified molecular-input line-entry system (SMILES) strings, circumventing the need for computationally heavy quantum-mechanical calculations.
The model builds on a 1.8-billion-parameter foundation encoder pre-trained on chemical compounds and subsequently fine-tuned on the QM8 benchmark dataset from MoleculeNet. It targets 12 distinct electronic transition scalar values computed through time-dependent density functional theory (TDDFT), including excitation energies and oscillator strengths.
This system serves computational chemists and molecular researchers studying optical or photoactive materials, providing fast evaluations of excited-state electronic properties for candidate molecules within synthetically accessible chemical space.
How it works
The model relies on a RoBERTa-PreLayerNorm encoder architecture featuring 24 layers, a hidden dimension of 2,304, an intermediate dimension of 9,216, 18 attention heads, and a context window of 2,048 tokens. The underlying foundation model was pre-trained using a masked language modelling objective on Kekulized SMILES strings from Enamine REAL Space. For fine-tuning on the QM8 dataset, the encoder is paired with a task-specific network consisting of a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Unit (GELU) activations and dropout.
Molecules are tokenised using the SMIRK tokeniser, typically taking Kekulized SMILES as input. During inference, sequence representations are extracted from the hidden state corresponding to the first token and passed into the regression head. The fine-tuning procedure optimised a multi-task mean squared error (MSE) loss using the AdamW optimiser to simultaneously predict all 12 target values.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder with a 2-layer MLP head
- Parameter count
- 1.8B
- Hidden dimension
- 2304
- Intermediate size
- 9216
- Attention heads
- 18
- Maximum sequence length
- 2048
- Pre-training data
- Enamine REAL Space
- Fine-tuning data
- QM8 dataset (21,786 small molecules)
- Base model
- mist-1.8B-dh61satt
What you can use it for
- Electronic Spectra Prediction
- Predicting 12 excited-state electronic transition metrics (eight excitation energies and four oscillator strengths) directly from molecular SMILES.
- Molecular Representation Extraction
- Generating pooled molecular embedding vectors from the first token representation of the pre-trained transformer encoder.
- Screening Photophysical Candidate Molecules
- Evaluating large collections of small molecules for desirable absorption or excitation profiles without running costly time-dependent density functional theory simulations.
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 1.8B · QM8
Strengths and limits
Good at
- Predicts 12 electronic transition properties simultaneously in a single forward pass.
- Pre-trained on a comprehensive collection of synthetically accessible compounds from Enamine REAL Space.
- Leverages a large 1.8-billion-parameter RoBERTa-PreLayerNorm backbone supporting sequence lengths up to 2,048 tokens.
Watch out for
- Predictions are restricted to the 12 electronic transition targets represented within the QM8 dataset.
- Requires input molecules to be formatted as SMILES strings, specifically expecting Kekulized SMILES representations.
- Using the SMIRK tokeniser introduces a software dependency on a local Rust installation.
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 1.8B · 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 1.8B · 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 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What properties does the model predict?
It predicts 12 scalar electronic spectra properties calculated via time-dependent density functional theory (TDDFT): eight excitation energies and four oscillator strengths.
What input format is required?
The model expects molecular SMILES strings, which should generally be Kekulized.
How are molecular embeddings pooled from the encoder?
Embeddings are extracted by pooling the encoder hidden state of the first token in the sequence.
What dataset was used to fine-tune this model?
It was fine-tuned on the QM8 benchmark dataset from MoleculeNet, which comprises 21,786 small organic molecules.
How to cite it
Wadell, Alexius and Bhutani, Anoushka and Azumah, Victor and Ellis-Mohr, Austin R. and Kelly, Celia and Zhao, Hancheng and Nayak, Anuj K. and Hegazy, Kareem and Brace, Alexander and Lin, Hongyi and Emani, Murali and Vishwanath, Venkatram and Gering, Kevin and Alkan, Melisa and Gibbs, Tom and Wells, Jack and Varshney, Lav R. and Ramsundar, Bharath and Duraisamy, Karthik and Mahoney, Michael W. and Ramanathan, Arvind and Viswanathan, Venkatasubramanian. Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 [physics], 2025. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 1.8B · 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.