Paper Link👁️

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

1. Introduction

Today, we’re introducing DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token. Compared with DeepSeek 67B, DeepSeek-V2 achieves stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times.

We pretrained DeepSeek-V2 on a diverse and high-quality corpus comprising 8.1 trillion tokens. This comprehensive pretraining was followed by a process of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unleash the model's capabilities. The evaluation results validate the effectiveness of our approach as DeepSeek-V2 achieves remarkable performance on both standard benchmarks and open-ended generation evaluation.

2. Model Downloads

Model	Context Length	Download
DeepSeek-V2	128k	🤗 HuggingFace
DeepSeek-V2-Chat(RL)	128k	🤗 HuggingFace

Due to the constraints of HuggingFace, the open-source code currently experiences slower performance than our internal codebase when running on GPUs with Huggingface. To facilitate the efficient execution of our model, we offer a dedicated vllm solution that optimizes performance for running our model effectively.

3. Evaluation Results

Base Model

Standard Benchmark

Benchmark	Domain	LLaMA3 70B	Mixtral 8x22B	DeepSeek V1 (Dense-67B)	DeepSeek V2 (MoE-236B)
MMLU	English	78.9	77.6	71.3	78.5
BBH	English	81.0	78.9	68.7	78.9
C-Eval	Chinese	67.5	58.6	66.1	81.7
CMMLU	Chinese	69.3	60.0	70.8	84.0
HumanEval	Code	52.4	39.0	42.7	40.9
MBPP	Code	68.6	64.2	57.4	66.6
GSM8K	Math	83.0	80.3	63.4	79.2
Math	Math	42.2	42.5	18.7	43.6

For more evaluation details, such as few-shot settings and prompts, please check our paper.

Context Window

Evaluation results on the Needle In A Haystack (NIAH) tests. DeepSeek-V2 performs well across all context window lengths up to 128K.

Chat Model

Standard Benchmark

Benchmark	Domain	QWen1.5 72B Chat	Mixtral 8x22B	LLaMA3 70B Instruct	DeepSeek V1 Chat (SFT)	DeepSeek V2 Chat(SFT)	DeepSeek V2 Chat(RL)
MMLU	English	76.2	77.8	80.3	71.1	78.4	77.8
BBH	English	65.9	78.4	80.1	71.7	81.3	79.7
C-Eval	Chinese	82.2	60.0	67.9	65.2	80.9	78.0
CMMLU	Chinese	82.9	61.0	70.7	67.8	82.4	81.6
HumanEval	Code	68.9	75.0	76.2	73.8	76.8	81.1
MBPP	Code	52.2	64.4	69.8	61.4	70.4	72.0
LiveCodeBench (0901-0401)	Code	18.8	25.0	30.5	18.3	28.7	32.5
GSM8K	Math	81.9	87.9	93.2	84.1	90.8	92.2
Math	Math	40.6	49.8	48.5	32.6	52.7	53.9

English Open Ended Generation Evaluation

We evaluate our model on AlpacaEval 2.0 and MTBench, showing the competitive performance of DeepSeek-V2-Chat-RL on English conversation generation.

Chinese Open Ended Generation Evaluation

Alignbench (https://arxiv.org/abs/2311.18743)

模型	开源/闭源	总分	中文推理	中文语言
gpt-4-1106-preview	闭源	8.01	7.73	8.29
DeepSeek-V2 Chat(RL)	开源	7.91	7.45	8.35
erniebot-4.0-202404(文心一言)	闭源	7.89	7.61	8.17
DeepSeek-V2 Chat(SFT)	开源	7.74	7.30	8.17
gpt-4-0613	闭源	7.53	7.47	7.59
erniebot-4.0-202312(文心一言)	闭源	7.36	6.84	7.88
moonshot-v1-32k-202404(月之暗面)	闭源	7.22	6.42	8.02
Qwen1.5-72B-Chat(通义千问)	开源	7.19	6.45	7.93
DeepSeek-67B-Chat	开源	6.43	5.75	7.11
Yi-34B-Chat(零一万物)	开源	6.12	4.86	7.38
gpt-3.5-turbo-0613	闭源	6.08	5.35	6.71

Coding Benchmarks

We evaluate our model on LiveCodeBench (0901-0401), a benchmark designed for live coding challenges. As illustrated, DeepSeek-V2 demonstrates considerable proficiency in LiveCodeBench, achieving a Pass@1 score that surpasses several other sophisticated models. This performance highlights the model's effectiveness in tackling live coding tasks.

4. Model Architecture

DeepSeek-V2 adopts innovative architectures to guarantee economical training and efficient inference：

For attention, we design MLA (Multi-head Latent Attention), which utilizes low-rank key-value union compression to eliminate the bottleneck of inference-time key-value cache, thus supporting efficient inference.
For Feed-Forward Networks (FFNs), we adopt DeepSeekMoE architecture, a high-performance MoE architecture that enables training stronger models at lower costs.

5. Chat Website

You can chat with the DeepSeek-V2 on DeepSeek's official website: chat.deepseek.com

6. API Platform

We also provide OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com. Sign up for over millions of free tokens. And you can also pay-as-you-go at an unbeatable price.

7. How to run locally

To utilize DeepSeek-V2 in BF16 format for inference, 80GB*8 GPUs are required.

Inference with Huggingface's Transformers

You can directly employ Huggingface's Transformers for model inference.

Text Completion

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "deepseek-ai/DeepSeek-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, device_map="auto", torch_dtype=torch.bfloat16, max_memory=max_memory)
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id

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 = "deepseek-ai/DeepSeek-V2-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, device_map="auto", torch_dtype=torch.bfloat16, max_memory=max_memory)
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id

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=100)

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

The complete chat template can be found within tokenizer_config.json located in the huggingface model repository.

An example of chat template is as belows:

<｜begin▁of▁sentence｜>User: {user_message_1}

Assistant: {assistant_message_1}<｜end▁of▁sentence｜>User: {user_message_2}

Assistant:

You can also add an optional system message:

<｜begin▁of▁sentence｜>{system_message}

User: {user_message_1}

Assistant: {assistant_message_1}<｜end▁of▁sentence｜>User: {user_message_2}

Assistant:

8. License

This code repository is licensed under the MIT License. The use of DeepSeek-V2 Base/Chat models is subject to the Model License. DeepSeek-V2 series (including Base and Chat) supports commercial use.

9. Citation

@misc{deepseek-v2,
  author = {DeepSeek-AI},
  title  = {DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model},
  year   = {2024},
  note   = {GitHub repository},
  url    = {https://github.com/deepseek-ai/deepseek-v2}
  }

10. Contact

If you have any questions, please raise an issue or contact us at [email protected].