LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling

By Yuxin Chen*, Chumeng Liang*, Hangke Sui*, Ruihan Guo, Chaoran Cheng, Jiaxuan You, Ge Liu.

The first continuous diffusion language model that rivals discrete counterparts on standard language modeling benchmarks like LM1B and OpenWebText.

TODO

• Inference code
• PPL evaluation code
• OpenWebText checkpoint on HuggingFace
• LM1B checkpoint on HuggingFace
• Training code (after paper acceptance)
• All trainable checkpoints (after paper acceptance)

Quick Start

1. Install dependencies

conda create -n langflow python=3.12
conda activate langflow
# Install CUDA-enabled torch first (adjust cu124 to match your driver)
pip install torch --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

2. Download the checkpoint

Download only the safetensors weights file from HuggingFace — no need to clone the HF repo:

# Using huggingface-hub CLI
hf download Continuous-Rivals-Discrete/langflow-owt model.safetensors --local-dir ./checkpoints

3. Run inference

python inference.py \
    --checkpoint ./checkpoints/model.safetensors \
    --num_samples 5 \
    --batch_size 1 \
    --num_steps 1024 \
    --seq_length 1024 \
    --seed 42 \
    --output samples.txt

Citation

@article{chen2026langflow,
  title={LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling},
  author={Chen, Yuxin and Liang, Chumeng and Sui, Hangke and Guo, Ruihan and Cheng, Chaoran and You, Jiaxuan and Liu, Ge},
  journal={arXiv preprint arXiv:2604.11748},
  year={2026}
}