DECO-0.2B

This is the 0.2B DECO checkpoint introduced by the paper DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices.

DECO (Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices) is a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. It is an improved version of the BlockFFN architecture.

• Authors: Chenyang Song, Weilin Zhao, Xu Han, Chaojun Xiao, Yingfa Chen, Zhiyuan Liu
• Paper: arXiv:2605.10933
• Code: https://github.com/thunlp/DECO

Quick start

You can load and use this model with AutoTokenizer and AutoModelForCausalLM from transformers. Since the model uses a custom architecture, trust_remote_code=True is required.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "SparseLLM/DECO-0.2B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).to("cuda").eval()

prompt = "Mixture-of-Experts models are useful because"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=64, do_sample=False)

print(tokenizer.decode(output[0], skip_special_tokens=True))

Citation

If you find our work useful for your research, please kindly cite our paper as follows:

@article{song2026deco,
      title={{DECO}: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices}, 
      author={Chenyang Song, Weilin Zhao, Xu Han, Chaojun Xiao, Yingfa Chen, Zhiyuan Liu},
      journal={arXiv preprint arXiv:2605.10933},
      year={2026},
      url={https://arxiv.org/pdf/2605.10933}, 
}