rednote-hilab / dots.llm1.inst

dots1

🤗 Hugging Face |    📑 Paper
🖥️ Demo |   💬 WeChat (微信) |   📕 rednote

Visit our Hugging Face (click links above), search checkpoints with names starting with dots.llm1 or visit the dots1 collection , and you will find all you need! Enjoy!

News

• 2025.06.06: We released the dots.llm1 series. Check our report for more details!

1. Introduction

The dots.llm1 model is a large-scale MoE model that activates 14B parameters out of a total of 142B parameters, delivering performance on par with state-of-the-art models. Leveraging our meticulously crafted and efficient data processing pipeline, dots.llm1 achieves performance comparable to Qwen2.5-72B after pretrained on high-quality corpus without synthetic data. To foster further research, we open-source intermediate training checkpoints spanning the entire training process, providing valuable insights into the learning dynamics of large language models.

2. Model Summary

This repo contains the base and instruction-tuned dots.llm1 model . which has the following features:

• Type: A MoE model with 14B activated and 142B total parameters trained on high-quality corpus.
• Training Stages: Pretraining and SFT.
• Architecture: Multi-head Attention with QK-Norm in attention Layer, fine-grained MoE utilizing top-6 out of 128 routed experts, plus 2 shared experts.
• Number of Layers: 62
• Number of Attention Heads: 32
• Supported Languages: English, Chinese
• Context Length: 32,768 tokens
• License: MIT

The highlights from dots.llm1 include:

• Enhanced Data Processing : We propose a scalable and fine-grained three-stage data processing framework designed to generate large-scale, high-quality and diverse data for pretraining.
• No Synthetic Data during Pretraining : High-quality non-synthetic tokens was used in base model pretraining.
• Performance and Cost Efficiency : dots.llm1 is an open-source model that activates only 14B parameters at inference, delivering both comprehensive capabilities and high computational efficiency.
• Infrastructure : We introduce an innovative MoE all-to-all communication and computation overlapping recipe based on interleaved 1F1B pipeline scheduling and an efficient grouped GEMM implementation to boost computational efficiency.
• Open Accessibility to Model Dynamics : Intermediate model checkpoints are released spanning the entire training process, facilitating future research into the learning dynamics of large language models.

3. Example Usage

Model Downloads

Docker (recommended)

The docker images are available on Docker Hub , based on the official images.

You can start a server via vllm.

docker run --gpus all \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    -p 8000:8000 \
    --ipc=host \
    rednotehilab/dots1:vllm-openai-v0.9.0.1 \
    --model rednote-hilab/dots.llm1.inst \
    --tensor-parallel-size 8 \
    --trust-remote-code \
    --served-model-name dots1

Then you can verify whether the model is running successfully in the following way.

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "dots1",
        "messages": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Who won the world series in 2020?"}
        ],
        "max_tokens": 32,
        "temperature": 0
    }'

Inference with huggingface

We are working to merge it into Transformers ( PR #38143 ).

Text Completion

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "rednote-hilab/dots.llm1.base"
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)

text = "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_new_tokens=100)
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Chat Completion

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "rednote-hilab/dots.llm1.inst"
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.bfloat16)

messages = [
    {"role": "user", "content": "Write a piece of quicksort code in C++"}
]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=200)

result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)

Inference with vllm

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Official support for this feature is covered in PR #18254 .

vllm serve dots.llm1.inst --port 8000 --tensor-parallel-size 8

An OpenAI-compatible API will be available at http://localhost:8000/v1 .

Inference with sglang

SGLang is a fast serving framework for large language models and vision language models. SGLang could be used to launch a server with OpenAI-compatible API service. Official support for this feature is covered in PR #6471 .

Getting started is as simple as running:

python -m sglang.launch_server --model-path dots.llm1.inst --tp 8 --host 0.0.0.0 --port 8000

An OpenAI-compatible API will be available at http://localhost:8000/v1 .

4. Evaluation Results

Detailed evaluation results are reported in this 📑 report .

Citation

If you find dots.llm1 is useful or want to use in your projects, please kindly cite our paper:

@misc{huo2025dotsllm1technicalreport,
      title={dots.llm1 Technical Report}, 
      author={Bi Huo and Bin Tu and Cheng Qin and Da Zheng and Debing Zhang and Dongjie Zhang and En Li and Fu Guo and Jian Yao and Jie Lou and Junfeng Tian and Li Hu and Ran Zhu and Shengdong Chen and Shuo Liu and Su Guang and Te Wo and Weijun Zhang and Xiaoming Shi and Xinxin Peng and Xing Wu and Yawen Liu and Yuqiu Ji and Ze Wen and Zhenhai Liu and Zichao Li and Zilong Liao},
      year={2025},
      eprint={2506.05767},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.05767}, 
}