---
license: mit
datasets:
  - ZINC-22
language:
  - en
tags:
  - molecular-generation
  - drug-discovery
  - llama
  - flash-attention
pipeline_tag: text-generation
---

# NovoMolGen

NovoMolGen is a family of molecular foundation models trained on 
1.5 billion ZINC-22 molecules with Llama architectures and FlashAttention. 
It achieves state-of-the-art performance on both unconstrained and 
goal-directed molecule generation tasks.

<img src="assets/NovoMolGen.png" width="900"/>

## How to load

```python
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> tokenizer = AutoTokenizer.from_pretrained("chandar-lab/NovoMolGen_32M_SMILES_BPE", trust_remote_code=True)
>>> model = AutoModelForCausalLM.from_pretrained("chandar-lab/NovoMolGen_32M_SMILES_BPE", trust_remote_code=True)
```

## Quick-start (FlashAttention + bf16)

```python
>>> from accelerate import Accelerator

>>> acc = Accelerator(mixed_precision='bf16')
>>> model = acc.prepare(model)

>>> outputs = model.sample(tokenizer=tokenizer, batch_size=4)
>>> print(outputs['SMILES'])
```

## Transformers-native HF checkpoint (`revision="hf-checkpoint"`)

We also publish a Transformers-native checkpoint on the `hf-checkpoint` revision. This version loads directly with `AutoModelForCausalLM` and works out-of-the-box with `.generate(...)`.

```python 
>>> import torch
>>> from transformers import AutoTokenizer, AutoModelForCausalLM

>>> model = AutoModelForCausalLM.from_pretrained("chandar-lab/NovoMolGen_32M_SMILES_BPE", revision='hf-checkpoint', device_map='auto')
>>> tokenizer = AutoTokenizer.from_pretrained("chandar-lab/NovoMolGen_32M_SMILES_BPE", revision='hf-checkpoint')

>>> input_ids = torch.tensor([[tokenizer.bos_token_id]]).expand(4, -1).contiguous().to(model.device)
>>> outs = model.generate(input_ids=input_ids, temperature=1.0, max_length=64, do_sample=True, pad_token_id=tokenizer.eos_token_id)

>>> molecules = [t.replace(" ", "") for t in tokenizer.batch_decode(outs, skip_special_tokens=True)]
['CCO[C@H](CNC(=O)N(CC(=O)OC(C)(C)C)c1cccc(Br)n1)C(F)(F)F',
'CCn1nnnc1CNc1ncnc(N[C@H]2CCO[C@@H](C)C2)c1C',
'CC(C)(O)CNC(=O)CC[C@H]1C[C@@H](NC(=O)COCC(F)F)C1',
'Cc1ncc(C(=O)N2C[C@H]3[C@H](CNC(=O)c4cnn[nH]4)CCC[C@H]3C2)n1C']
```

## Citation

```bibtex
@misc{chitsaz2025novomolgenrethinkingmolecularlanguage,
      title={NovoMolGen: Rethinking Molecular Language Model Pretraining}, 
      author={Kamran Chitsaz and Roshan Balaji and Quentin Fournier and Nirav Pravinbhai Bhatt and Sarath Chandar},
      year={2025},
      eprint={2508.13408},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.13408}, 
}
```