Technical Blog | MiniCPM Paper | MiniCPM-V Repo | Join our discord and WeChat

MiniCPM is an End-Side LLM developed by ModelBest Inc. and TsinghuaNLP, with only 2.4B parameters excluding embeddings (2.7B in total).

• MiniCPM has very close performance compared with Mistral-7B on open-sourced general benchmarks with better ability on Chinese, Mathematics and Coding after SFT. The overall performance exceeds Llama2-13B, MPT-30B, Falcon-40B, etc.
• After DPO, MiniCPM outperforms Llama2-70B-Chat, Vicuna-33B, Mistral-7B-Instruct-v0.1, Zephyr-7B-alpha, etc. on MTBench.
• MiniCPM-V 2.0, based on MiniCPM-2B, achieves state-of-the-art performance on multiple benchmarks among models under 7B parameters. It even outperforms strong Qwen-VL-Chat 9.6B, CogVLM-Chat 17.4B, and Yi-VL 34B on OpenCompass. MiniCPM-V 2.0 also shows strong OCR capability, achieving comparable performance to Gemini Pro in scene-text understanding.
• MiniCPM can be deployed and infer on smartphones, and the speed of streaming output is relatively higher than human verbal speed. MiniCPM-V has also successfully deployed multi-modal models on smartphones.
• The cost of developing based on MiniCPM is low. Parameter efficient finetuning can be conducted with a single 1080/2080 GPU and full parameter finetuning can be conducted with a 3090/4090 GPU.

We release all model parameters for research and limited commercial use.

• SFT and DPO version based on MiniCPM-2B: MiniCPM-2B-SFT/DPO
• The multi-modal model MiniCPM-V 2.0 based on MiniCPM-2B.
• The INT4 quantized version MiniCPM-2B-SFT/DPO-Int4 based on MiniCPM-2B-SFT/DPO
• The 128k long context version of MiniCPM-2B: MiniCPM-2B-128k.
• The MoE version of MiniCPM-2B: MiniCPM-MoE-8x2B.
• SFT version of MiniCPM-1B, a lighter-weight model: MiniCPM-1B-SFT.
• Mobile phone application based on MLC-LLM and LLMFarm. Both language model and multimodel model can conduct inference on smartphones.
• 30 Intermidiate checkpoints of MiniCPM-2B for academic purpose.

• Due to limitations in model size, the model may experience hallucinatory issues. As DPO model tend to generate longer response, hallucinations are more likely to occur. We will also continue to iterate and improve the MiniCPM model.
• To ensure the generality of the model for academic research purposes, we have not subject it to any identity-specific training. Meanwhile, as we use ShareGPT open-source corpus as part of the training data, the model may output identity-related information similar to the GPT series models.
• Due to the limitation of model size, the output of the model is greatly influenced by prompts, which may result in inconsistent results from multiple attempts.
• Due to limited model capacity, the model's knowledge recall may not be accurate. In the future, we will combine the RAG method to enhance the model's knowledge retention ability.

• Updates
• Downloading
• Quick Start
• Community
• Benchmark
• Deployment on Mobile Phones
• Demo & API
• Fine-tuning Models
• LICENSE
• Citation
• Show Cases

• 2024/04/11 We release MiniCPM-V 2.0, MiniCPM-2B-128k, MiniCPM-MoE-8x2B and MiniCPM-1B! Click here to read our technical blog.
• 2024/03/16 Intermediate checkpoints were released here!
• 2024/02/13 We support llama.cpp
• 2024/02/09 We have included a Community section in the README to encourage support for MiniCPM from the open-source community.
• 2024/02/08 We updated the llama-format model weights, which can be loaded into LlamaModel directly, making it more convenient for everyone to use our model quickly.
• 2024/02/01 Initial release.

• Language Model

  HuggingFace	ModelScope	WiseModel
  MiniCPM-2B-sft-bf16	MiniCPM-2B-sft-bf16	MiniCPM-2B-sft-bf16
  MiniCPM-2B-dpo-bf16	MiniCPM-2B-dpo-bf16	MiniCPM-2B-dpo-bf16
  MiniCPM-2B-128k	MiniCPM-2B-128k
  MiniCPM-MoE-8x2B	MiniCPM-MoE-8x2B
  MiniCPM-1B-sft-bf16	MiniCPM-1B-sft-bf16

  Note: More model versions can be found here.

• Multimodel Model

  HuggingFace	ModelScope	WiseModel
  MiniCPM-V 2.0	MiniCPM-V 2.0
  MiniCPM-V	MiniCPM-V	MiniCPM-V
  OmniLMM-12B	OmniLMM-12B	OmniLMM-12B

• Colab

• Install transformers>=4.36.0 and accelerate，run the following python code.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
torch.manual_seed(0)

path = 'openbmb/MiniCPM-2B-dpo-bf16'
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map='cuda', trust_remote_code=True)

responds, history = model.chat(tokenizer, "Which city is the capital of China?", temperature=0.8, top_p=0.8)
print(responds)

• Examples

The capital city of China is Beijing. Beijing is not only the political center of China but also a cultural and economic hub. It is known for its rich history and numerous landmarks, such as the Great Wall, the Forbidden City, and the Temple of Heaven. The city is also home to the National Stadium, also known as the "Bird's Nest," and the National Aquatics Center, or "Water Cube." Beijing is a significant city in China, with a population of over 21 million people.

To facilitate ease of use, we have converted the model weights of MiniCPM to adapt to the structure of the LLaMA model:

import torch
from transformers import LlamaTokenizerFast, LlamaForCausalLM
model_path = "openbmb/MiniCPM-2B-dpo-bf16-llama-format"
tokenizer = LlamaTokenizerFast.from_pretrained(model_path)
model = LlamaForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map='cuda', trust_remote_code=True)

prompt="Now you act like a terminal situated within a beginner's C   practice repository folder, please provide the output for the command: `ls -l`"
input_ids = tokenizer.encode("<User>{}<AI>".format(prompt), return_tensors='pt', add_special_tokens=True).cuda()
responses = model.generate(input_ids, temperature=0.3, top_p=0.8, repetition_penalty=1.02, max_length=1024)
responses = tokenizer.decode(responses[0], skip_special_tokens=True)
print(responses)

import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained('openbmb/MiniCPM-V', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V', trust_remote_code=True)
model.eval().cuda()

image = Image.open('xx.jpg').convert('RGB')
question = 'What is in the image?'
msgs = [{'role': 'user', 'content': question}]

res, context, _ = model.chat(
    image=image,
    msgs=msgs,
    context=None,
    tokenizer=tokenizer,
    sampling=True,
    temperature=0.7
)
print(res)

• Install vLLM main: build from source。

• Examples

  python inference/inference_vllm.py --model_path <hf_repo_path> --prompt_path prompts/prompt_demo.txt

We have supported inference with llama.cpp, ollama, fastllm, mlx_lm. Thanks to @runfuture for the adaptation of llama.cpp and ollama.

llama.cpp

1. install llama.cpp
2. download model in gguf format. link-fp16 link-q4km
3. In command line:

./main -m ../../model_ckpts/download_from_hf/MiniCPM-2B-dpo-fp16-gguf.gguf --prompt "<用户>Write an acrostic poem with the word MINICPM (One line per letter)<AI>" --temp 0.3 --top-p 0.8 --repeat-penalty 1.05

More parameters adjustment see this

ollama Solving this issue

fastllm

1. install fastllm
2. inference

import torch
from transformers import AutoTokenizer, LlamaTokenizerFast, AutoModelForCausalLM
path = 'openbmb/MiniCPM-2B-dpo-fp16'
tokenizer = AutoTokenizer.from_pretrained(path)
model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.float16, device_map='cuda', trust_remote_code=True)
from fastllm_pytools import llm
llm.set_device_map("cpu")
model = llm.from_hf(model, tokenizer, dtype = "float16") # dtype支持 "float16", "int8", "int4"
print(model.response("<用户>Write an acrostic poem with the word MINICPM (One line per letter)<AI>", top_p=0.8, temperature=0.5, repeat_penalty=1.02))

mlx_lm

1. install mlx_lm

   pip install mlx_lm

2. download model weights MiniCPM-2B-sft-bf16-llama-format-mlx

3. inference

   python -m mlx_lm.generate --model mlx-community/MiniCPM-2B-sft-bf16-llama-format-mlx --prompt "hello, tell me a joke." --trust-remote-code

• ChatLLM: Run MiniCPM on CPU

• Since it is difficult to standardize the evaluation of LLMs and there is no public prompt and test code for a large number of evaluations, we can only try our best to make it suitable for all types of models in terms of specific evaluation methods.
• Overall, we use a unified prompt input for testing, and adjust the input according to the corresponding template for each model.
• The evaluation scripts and prompts have been open-sourced in our Github repository, and we welcome more developers to continuously improve our evaluation methods.
  • For the text evaluation part, we use our open source large model capability evaluation framework UltraEval. The following is the open source model reproduction process:

    • install UltraEval

      git clone https://github.com/OpenBMB/UltraEval.git
      cd UltraEval
      pip install -e .

    • Download the relevant data and unzip it for processing

      wget -O RawData.zip "https://cloud.tsinghua.edu.cn/f/71b5232264ae4833a4d0/?dl=1"
      unzip RawData.zip
      python data_process.py

    • Execute evaluation scripts (templates are provided and can be customized)

      bash run_eval.sh

• Because MiniCPM uses the structure of Mup, which is slightly different from existing models in terms of specific computations, we have based the implementation of our model on the vllm=0.2.2 version.
• For non-MiniCPM models, we directly sampled the latest version of vllm=0.2.7 for inference.

• For the QA task (multiple-choice task), we chose to test in two ways:
  • PPL: The options are used as a continuation of the question generation and the answer selection is based on the PPL of each option;
  • The second is to generate the answer options directly.
• For different models, the results obtained by these two approaches vary widely. the results on both MiniCPM models are closer, while models such as Mistral-7B-v0.1 perform better on PPL and worse on direct generation.
• In the specific evaluation, we take the higher score of the two evaluation methods as the final result, so as to ensure the fairness of the comparison (* in the following table indicates the PPL).

Note：* means evaluation results are directly taken from their technical reports. † means evaluation results on the full set of MBPP, instead of the hand-verified set.

• After INT4 quantization, MiniCPM only occupies 2GB of space, meeting the requirements of inference on end devices.
• We have made different adaptations for different operating systems.
• Note: The current open-source framework is still improving its support for mobile phones, and not all chips and operating system versions can successfully run MLC-LLM or LLMFarm.
• Android, HarmonyOS
  • Adapt based on open-source framework MLC-LLM.
  • Adapted for text model MiniCPM, and multimodel model MiniCPM-V.
  • Support MiniCPM-2B-SFT-INT4, MiniCPM-2B-DPO-INT4, and MiniCPM-V.
  • Compile and Installation Guide
• iOS
  • Adapt based on open-source framework LLMFarm.
  • Adapted for text model MiniCPM.
  • Support MiniCPM-2B-SFT-INT4, MiniCPM-2B-DPO-INT4.
  • Compile and Installation Guide

• We did not conduct in-depth optimization and system testing on the mobile inference model, only verifying the feasibility of MiniCPM using mobile phone chips for inference. We welcome more developers to continuously improve the inference performance of LLMs on mobile phones and update the test results below.

• We have also verified the feasibility of deploying MiniCPM-V series models on mobile phones based on MLC-LLM, and it can input and output normally. However, there also exist a problem of long image processing time, which needs further optimization. The demo video below is the raw screen recording on a Xiaomi 14 Pro without edition.

Using the following command can launch the gradio-based demo.

# generation powered by vllm
python demo/vllm_based_demo.py --model_path <vllmcpm_repo_path>
# generation powered by huggingface
python demo/hf_based_demo.py --model_path <hf_repo_path>

• Parameter-efficient Tuning

  • With parameter-efficient tuning, we can tune MiniCPM using one piece of NVIDIA GeForce GTX 1080/2080.
  • Code for Parameter-efficient Tuning

• Full-parameter Tuning

  • Using BMTrain，as well as checkpointing and ZeRO-3 (zero redundancy optimizer)，we can tune all parameters of MiniCPM using one piece of NVIDIA GeForce GTX 3090/4090.
  • This code will be available soon.

• mlx Parameter-efficient Tuning

  • environment preparation

    pip install -r finetune/requirements_mlx.txt

  • finetune

    # train
    python mlx_finetune.py --model MiniCPM-2B-sft-bf16-llama-format-mlx  --data data/AdvertiseGen  --train  --seed 2024 --iters 500
    # test
    python mlx_finetune.py --model MiniCPM-2B-sft-bf16-llama-format-mlx  --data data/AdvertiseGen  --test --seed 2024

• This repository is released under the Apache-2.0 License.
• The usage of MiniCPM model weights must strictly follow the General Model License (GML).
• The models and weights of MiniCPM are completely free for academic research.
• If you intend to utilize the model for commercial purposes, please reach out to [email protected] to obtain the certificate of authorization.

• As a language model, MiniCPM generates content by learning from a vast amount of text.
• However, it does not possess the ability to comprehend or express personal opinions or value judgments.
• Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers.
• Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.

• Please cite our paper if you find our work valuable.

@misc{hu2024minicpm,
      title={MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies}, 
      author={Shengding Hu and Yuge Tu and Xu Han and Chaoqun He and Ganqu Cui and Xiang Long and Zhi Zheng and Yewei Fang and Yuxiang Huang and Weilin Zhao and Xinrong Zhang and Zheng Leng Thai and Kaihuo Zhang and Chongyi Wang and Yuan Yao and Chenyang Zhao and Jie Zhou and Jie Cai and Zhongwu Zhai and Ning Ding and Chao Jia and Guoyang Zeng and Dahai Li and Zhiyuan Liu and Maosong Sun},
      year={2024},
      eprint={2404.06395},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}