Top 10 AI Model Created by Chinese Company

Table of Contents

What is LLM model？

Definition and overview

An AI model is a program that has been trained on a set of data to recognize certain patterns or make certain decisions without further human intervention.

Large language models, also known as LLMs, are very large deep learning models that are pre-trained on vast amounts of data.

The underlying transformer is a set of neural networks that consist of an encoder and a decoder with self-attention capabilities. The encoder and decoder extract meanings from a sequence of text and understand the relationships between words and phrases in it.

Which one is the best model for you?

AI large models are developing very rapidly. Different companies and research institutions present new research achievements daily, along with new large language models.

Therefore, we cannot definitively tell you which one is the best.

However, there are top-tier companies and models, such as OpenAI. There is now a set of standards and test questions to evaluate models.

You can refer to superclueai to view the model’s scores in various tasks and choose the one that suits you. Also, you can follow the latest news to know more about the ability of the LLM model.

Hunyuan-Large by Tencent

Model Introduction

On November 5th, Tencent releases Open-Source MoE Large Language Model Hunyuan-large with a total of 398 billion parameters, making it the largest in the industry, with 52 billion activation parameters.

Public evaluation results show that Tencent‘s Hunyuan Large model leads comprehensively in various projects.

Technical Advantages

High-Quality Synthetic Data: By enhancing training with synthetic data, Hunyuan-Large can learn richer representations, handle long-context inputs, and generalize better to unseen data.
KV Cache Compression: Utilizes Grouped Query Attention (GQA) and Cross-Layer Attention (CLA) strategies to significantly reduce memory usage and computational overhead of KV caches, improving inference throughput.
Expert-Specific Learning Rate Scaling: Sets different learning rates for different experts to ensure each sub-model effectively learns from the data and contributes to overall performance.
Long-Context Processing Capability: The pre-trained model supports text sequences up to 256K, and the Instruct model supports up to 128K, significantly enhancing the ability to handle long-context tasks.
Extensive Benchmarking: Conducts extensive experiments across various languages and tasks to validate the practical effectiveness and safety of Hunyuan-Large.

Inference Framework and Training Framework

This open-source release offers two inference backend options tailored for the Hunyuan-Large model: the popular vLLM-backend and the TensorRT-LLM Backend. Both solutions include optimizations for enhanced performance.

The Hunyuan-Large open-source model is fully compatible with the Hugging Face format, enabling researchers and developers to perform model fine-tuning using the hf-deepspeed framework. Additionally, we support training acceleration through the use of flash attention.

How to further use this model

This is an open-source model. You can find “tencent-hunyuan” on GitHub, where they provide detailed instructions and usage guides. You can further explore and research it to create more possibilities.

Moonshot(Kimi) by Moonshot AI

Summary Introduction

Moonshot is a large-scale language model developed by Dark Side of the Moon. Here is an overview of its features:

Technological Breakthrough: Moonshot achieves remarkable advancements in long-text processing, with its smart assistant product, Kimichat, supporting up to 2 million Chinese characters in lossless context input.
Model Architecture: By employing an innovative network structure and engineering optimizations, it achieves long-range attention without relying on “shortcut” solutions like sliding windows, downsampling, or smaller models that often degrade performance. This enables comprehensive understanding of ultra-long texts even with hundreds of billions of parameters.
Application-Oriented: Developed with a focus on practical application, Moonshot aims to become an indispensable daily tool for users, evolving based on real user feedback to generate tangible value.

Key Features

Long-Text Processing Ability: Capable of handling extensive texts such as novels or complete financial reports, offering users in-depth, comprehensive insights and summaries of long documents.
Multimodal Fusion: Integrates multiple modalities, combining text with image data to enhance analysis and generation capabilities.
High Language Understanding and Generation Capability: Demonstrates excellent multilingual performance, accurately interpreting user input and generating high-quality, coherent, and semantically appropriate responses.
Flexible Scalability: Offers strong scalability, allowing for customization and optimization based on different application scenarios and needs, providing developers and enterprises with significant flexibility and autonomy.

Usage Methods

API Integration: Users can register for an account on the Dark Side of the Moon official platform, apply for an API key, and then integrate Moonshot’s capabilities into their applications using the API with compatible programming languages.
Using Official Products and Tools: Directly use Kimichat, the smart assistant product based on the Moonshot model, or leverage associated tools and platforms offered by Dark Side of the Moon.
Integration with Other Frameworks and Tools: Moonshot can be integrated with popular AI development frameworks like LangChain to build more robust language model applications.

GLM-4-Plus by zhipu.ai

Summary Introduction

GLM-4-Plus, developed by Zhipu AI, is the latest iteration of the fully self-developed GLM foundation model, with significant enhancements in language comprehension, instruction-following, and long-text processing.

Key Features and Advantages

Strong Language Understanding: Trained on extensive datasets and optimized algorithms, GLM-4-Plus excels at handling complex semantics, accurately interpreting the meaning and context of various texts.
Outstanding Long-Text Processing: With an innovative memory mechanism and segmented processing technique, GLM-4-Plus can effectively handle long texts up to 128k tokens, making it highly proficient in data processing and information extraction.
Enhanced Reasoning Capabilities: Incorporates Proximal Policy Optimization (PPO) to maintain stability and efficiency while exploring optimal solutions, significantly improving the model’s performance in complex reasoning tasks like mathematics and programming.
High Instruction-Following Accuracy: Accurately understands and adheres to user instructions, generating high-quality, expectation-aligned text based on user requirements.

Usage Instructions

Register an Account and Obtain an API Key: First, register an account on Zhipu’s official website and acquire an API key.
Review Official Documentation: Refer to the official GLM-4 series documentation for detailed parameters and usage instructions.

SenseChat 5.5 by SenceTime

Summary Introduction

SenseChat 5.5, developed by SenseTime, is the 5.5 version of its large language model, based on the InternLM-123b, one of China’s earliest large language models built on trillions of parameters and continuously updated.

Key Features and Advantages

Powerful Comprehensive Performance: Consistently ranks among the top tier in a variety of evaluation tasks, excelling across fundamental competencies in humanities and sciences as well as advanced “Hard” tasks. It demonstrates superior performance in language understanding and security in humanities, and excels in logic and coding in sciences.
Efficient Edge Applications: SenseTime has released the SenseChat Lite-5.5 version, which reduces initial load time to just 0.19 seconds, a 40% improvement over SenseChat Lite-5.0 released in April, with inference speed reaching 90.2 characters per second and an annual cost per device as low as 9.9 yuan.
Exceptional Language Capabilities: As a natural language application, it effectively handles extensive text data, demonstrating robust natural language dialogue, logical reasoning abilities, broad knowledge, and frequent updates. It supports Simplified Chinese, Traditional Chinese, English, and common programming languages.

Usage and Application Products

Direct Use: Users can register on the [SenseTime website] to access SenseChat through the web or mobile app and interact with the model.
API Integration: SenseTime offers API access for businesses and developers, enabling them to integrate SenseChat 5.5 into their products or applications.

Qwen2.5-72B-Instruct by Qwen team, Alibaba Cloud

Model Inturduction

Qwen2.5 is the latest series of Qwen large language models. For Qwen2.5, the team released a number of base language models and instruction-tuned language models ranging from 0.5 to 72 billion parameters.

Key features

Dense, easy-to-use, decoder-only language models, available in 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B sizes, and base and instruct variants.
Pretrained on our latest large-scale dataset, encompassing up to 18T tokens.
Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g, tables), and generating structured outputs, especially JSON.
More resilient to the diversity of system prompts, enhancing role-play implementation and condition-setting for chatbots.
Context length supports up to 128K tokens and can generate up to 8K tokens.
Multilingual support for over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more.

How to quickly start?

You can find tutorials for using large models on Github and Hugging face. Based on these tutorials, you can effectively run the model and realize your functions and ideas.

Doubao-pro by Doubao Team, ByteDance

Summary Introduction

Doubao-pro is a large language model independently developed by ByteDance, officially released on May 15, 2024. In the Flageval evaluation platform for large models, Doubao-pro ranked second among closed-source models with a score of 75.96.

Versions: Doubao-pro includes versions with 4k, 32k, and 128k context windows, each supporting different context lengths for inference and fine-tuning.
Performance Improvement: According to ByteDance’s internal testing, Doubao-pro-4k achieved a total score of 76.8 across 11 industry-standard public benchmarks.

Key Features and Advantages

Strong Comprehensive Abilities: Doubao-pro excels in math, knowledge application, and problem-solving across objective and subjective evaluations.
Wide Range of Applications: As one of the most widely used and versatile domestic models, Doubao’s AI assistant, “Doubao,” ranks first in downloads among AIGC applications on the Apple App Store and major Android app markets.
High Cost-Effectiveness: Doubao-pro-32k’s inference input cost is only 0.0008 yuan per thousand tokens. For example, processing the Chinese version of Harry Potter (2.74 million characters) costs only 1.5 yuan.
Outstanding Language Understanding and Generation: Doubao-pro accurately comprehends diverse natural language inputs and generates high-quality, coherent, and logical responses, meeting user needs in simple Q&A, complex text creation, and explanations in specialized fields.
Efficient Inference Speed: With extensive data training and optimization, Doubao-pro offers an inference speed advantage, allowing for quick response times and enhanced user experience, especially when handling large volumes of text or complex tasks.

Usage Methods

Through Volcano Engine: Use Doubao-pro by calling the model’s API, with code samples available in the Volcano Engine’s official documentation.
For Specific Products: Doubao-pro is available to the enterprise market through the Volcano Engine, allowing businesses to integrate it into their products or services. You can also experience the Doubao model through the Doubao app.

360gpt2-pro by 360

Summary Introduction

Model Name: 360GPT2-Pro is part of the 360 Zhibrain large model series developed by 360.
Technical Foundation: Leveraging 20 years of security data, 10 years of AI experience, and the expertise of 80 AI and 100 security experts, 360 used 5,000 GPU resources over 200 days to train and optimize the Zhibrain model, with 360GPT2-Pro being one of its advanced versions.

Key Features and Advantages

Strong Language Generation: Excels in language generation tasks, especially in the humanities, by creating high-quality, creative, and logically coherent content, such as stories and copywriting.
Robust Knowledge Understanding and Application: Equipped with a broad knowledge base, it accurately interprets and applies information to answer questions and solve problems effectively.
Enhanced Retrieval-Based Generation: Competent in retrieval-augmented generation, particularly for Chinese, enabling the model to generate responses that are aligned with user needs and real-world data, reducing hallucination probability.
Enhanced Security Features: Benefiting from 360’s longstanding expertise in security, 360GPT2-Pro provides a level of safety and reliability, effectively addressing various security risks.

360AI Search: Integrates 360GPT2-Pro with search functionality to provide users with a more comprehensive and in-depth search experience.
360AI Browser: Incorporates 360GPT2-Pro into the 360AI Browser, allowing users to interact with the model via specific interfaces or through voice input to obtain information and suggestions.

Step-2-16k by stepfun

Summary Introduction

Developer: StepStar released the official version of the STEP-2 trillion-parameter language model in 2024, with step-2-16k referring to its variant supporting a 16k context window.
Model Architecture: Built on an innovative MoE (Mixture of Experts) architecture, which dynamically activates different expert models based on tasks and data distribution, enhancing both performance and efficiency.
Parameter Scale: With a trillion parameters, the model captures extensive language knowledge and semantic information, displaying powerful capabilities across various natural language processing tasks.

Key Features and Advantages

Powerful Language Understanding and Generation: Accurately interprets input text and generates high-quality, natural responses, supporting tasks such as answering questions, content generation, and conversational exchange with accuracy and value.
Multi-domain Knowledge Coverage: Trained on massive datasets, the model encompasses broad knowledge in areas such as mathematics, logic, programming, knowledge, and creative writing, making it versatile for cross-domain responses and applications.
Long Sequence Processing Capability: With a 16k context window, the model excels at handling long text sequences, facilitating comprehension and processing of lengthy articles and complex documents.
Performance Close to GPT-4: Achieving near-GPT-4 performance in multiple language tasks, this model showcases high-level comprehensive language processing abilities.

Usage and Applications

StepStar provides an open platform for enterprises and developers to apply for access to the step-2-16k model.

Users can integrate the model into applications or development projects through API calls, using platform-provided documentation and development tools to implement various natural language processing functionalities.

DeepSeek-V2.5 by deepseek

Summary Introduction

DeepSeek-V2.5, developed by the DeepSeek team, is a powerful open-source language model that integrates the capabilities of DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct, representing the culmination of previous model advancements. Key details are as follows:

Development History: In September 2024, they officially released DeepSeek-V2.5, combining chat and coding capabilities. This version enhances both general language proficiency and coding functionality.
Open Source Nature: In line with a commitment to open-source development, DeepSeek-V2.5 is now available on Hugging Face, allowing developers to adjust and optimize the model as needed.

Key Features and Advantages

Combined Language and Coding Abilities: DeepSeek-V2.5 retains the conversational abilities of a chat model and the coding strengths of a coder model, making it a true “all-in-one” solution capable of handling everyday conversations, complex instruction following, code generation, and completion.
Human Preference Alignment: Fine-tuned to align with human preferences, the model has been optimized for writing quality and instruction adherence, performing more naturally and intelligently across multiple tasks to better understand and meet user needs.
Outstanding Performance: DeepSeek-V2.5 surpasses previous versions on various benchmarks, and achieves top results in coding benchmarks like humaneval python and live code bench, showcasing its strength in instruction adherence and code generation.
Extended Context Support: With a maximum context length of 128k tokens, DeepSeek-V2.5 effectively handles long-form texts and multi-turn dialogues.
High Cost-Effectiveness: Compared to top-tier closed-source models like Claude 3.5 Sonnet and GPT-4o, DeepSeek-V2.5 offers a significant cost advantage.

Usage Methods

Via Web Platform: Access DeepSeek-V2.5 through web platforms like SiliconCloud’s DeepSeek-V2.5 playground.
Via API: Users can create an account to obtain an API key, then integrate DeepSeek-V2.5 into their systems through the API for secondary development and applications.
Local Deployment: Requires 8 GPUs at 80GB each, using Hugging Face’s Transformers for inference. Refer to documentation and sample code for specific steps.
Within Specific Products:
- Cursor: This AI code editor, based on VSCode, allows users to configure the DeepSeek-V2.5 model, connecting to SiliconCloud’s API for on-page code generation via shortcuts, enhancing coding efficiency.
- Other Development Tools or Platforms: Any development tool or platform that supports external language model APIs can theoretically integrate DeepSeek-V2.5 by obtaining an API key, enabling language generation and code writing capabilities.

Ernie-4.0-turbo-8k-preview by Baidu

Summary Introduction

Ernie-4.0-turbo-8k-preview is part of Baidu’s ERNIE 4.0 Turbo series, officially released on June 28, 2024, and fully opened to enterprise clients on July 5, 2024.

Key Features and Advantages

Performance Improvement: As an upgraded version of ERNIE 4.0, this model extends context input length from 2k tokens to 8k tokens, enabling it to handle larger datasets, read more documents or URLs, and perform better on tasks involving long texts.
Cost Reduction: The input and output costs of ERNIE 4.0-turbo-8k-preview are as low as 0.03 CNY per 1,000 tokens and 0.06 CNY per 1,000 tokens, a 70% price reduction from the general version of ERNIE 4.0.
Technical Optimization: Enhanced by turbo technology, this model achieves dual improvements in training speed and performance, allowing for faster model training and deployment.
Wide Application: Due to its performance and cost advantages, the model is widely applicable across fields such as intelligent customer service, virtual assistants, education, and entertainment, providing a smooth and natural conversation experience. Its robust generation capabilities also make it highly suitable for content creation and data analysis.

Usage

The ERNIE 4.0-turbo-8k-preview is primarily available to enterprise clients, who can access it via Baidu’s Qianfan Large Model Platform on Baidu Intelligent Cloud.