Introduction

Dream-Coder 7B is a diffusion LLM for code trained exclusively on open-source data across its development stages—adaptation, supervised fine-tuning, and reinforcement learning. It achieves an impressive 21.4% pass@1 on LiveCodeBench (2410-2505), outperforming other open-source diffusion LLMs by a wide margin.

News

• Sep 25, 2025: Released data processing, training, and evaluation scripts for the instruct model. See instruct.
• Sep 21, 2025: Released data details and evaluation scripts for the base model. See base.
• Sep 1, 2025: Our technical report was out.
• July 23, 2025: Try our online demo via HF space!
• July 15, 2025: Released Dream-Coder checkpoints, along with our blog post and Notion page.

Features

Flexible Code Generation

We observe Dream-Coder 7B exhibits emergent any-order generation that adaptively determines its decoding style based on the coding task. For example, Dream-Coder 7B Instruct displays patterns such as:

Sketch-First Generation	Left-to-Right Generation
	Iterative Back-and-Forth Generation

These demos were collected using consistent sampling parameters: temperature=0.1, diffusion_steps=512, max_new_tokens=512, alg="entropy", top_p=1.0, alg_temp=0.0, eos_penalty=3.0.

Variable-Length Code Infilling

We also introduce an infilling variant, DreamOn-7B, that naturally adjusts the length of masked spans during generation for variable-length code infilling. For more details, please refer to our accompanying blog post.

Quickstart

To get start with, please install transformers==4.46.2 and torch==2.5.1. Here is an example to use Dream-Coder 7B:

import torch
from transformers import AutoModel, AutoTokenizer

model_path = "Dream-org/Dream-Coder-v0-Instruct-7B"
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.to("cuda").eval()

messages = [
    {"role": "user", "content": "Write a quick sort algorithm."}
]
inputs = tokenizer.apply_chat_template(
    messages, return_tensors="pt", return_dict=True, add_generation_prompt=True
)
input_ids = inputs.input_ids.to(device="cuda")
attention_mask = inputs.attention_mask.to(device="cuda")

output = model.diffusion_generate(
    input_ids,
    attention_mask=attention_mask,
    max_new_tokens=768,
    output_history=True,
    return_dict_in_generate=True,
    steps=768,
    temperature=0.1,
    top_p=0.95,
    alg="entropy",
    alg_temp=0.,
)
generations = [
    tokenizer.decode(g[len(p) :].tolist())
    for p, g in zip(input_ids, output.sequences)
]

print(generations[0].split(tokenizer.eos_token)[0])

Acknowledgement

We gratefully acknowledge the following open-source projects, which have been instrumental to this work:

• verl: Reinforcement learning training framework
• SandboxFusion: Secure code execution environment
• Fast-dLLM: Inference acceleration for dLLMs

Citation

@article{xie2025dream,
  title={Dream-coder 7b: An open diffusion language model for code},
  author={Xie, Zhihui and Ye, Jiacheng and Zheng, Lin and Gao, Jiahui and Dong, Jingwei and Wu, Zirui and Zhao, Xueliang and Gong, Shansan and Jiang, Xin and Li, Zhenguo and others},
  journal={arXiv preprint arXiv:2509.01142},
  year={2025}
}