Large language Models

DeepSeek

Model name: DeepSeek-V3, DeepSeek-R1
Development organization: DeepSeek
Features:

• Open source model: Different from the closed source model abroad, it adopts an open source strategy, triggering a wave of reproduction at home and abroad, with a cost of only 6 million US dollars, far lower than international giants.
• High performance: It is comparable to OpenAI o1 in mathematics, code, and natural language reasoning tasks, and performs well in liberal arts tasks (total score 68.3, liberal arts 78.2).
• Algorithm innovation: Improve reasoning ability by optimizing training strategies and reduce dependence on expensive labeled data.

Qwen

Model name: Qwen2-72B, Qwen-max-latest, etc.
Development organization: Alibaba
Features:

• International competitiveness: Qwen2-72B surpassed Meta's Llama3-70B in the OpenCompass evaluation and became the most downloaded Chinese open source model in the world.
• Multimodal capability: Qwen-VL performed well in multimodal evaluation and supports image and text interaction.
• High cost performance: Reduce costs by optimizing the architecture, and the price has dropped to less than 0.5 yuan/million tokens

EnseChat

Model name: SenseChat 5.5-latest
Developer: SenseTime
Features:

• Liberal arts advantage: Liberal arts score 81.8, surpassing most international models, good at natural language generation and understanding.
• Industry application: Widely used in smart cities, medical and other fields, supporting semantic analysis of complex scenarios.

ERNIE

Model name: ERNIE-4.0-Turbo, Wenxinyiyan
Development organization: Baidu
Features:

• Chinese understanding: Outstanding performance in Chinese semantic processing, integrated into Baidu search, map and other products.
• Vertical field optimization: Specialized versions are launched for medical, education and other scenarios, supporting localized deployment.

Spark Cognitive Model

Development organization: iFlytek
Features:

• Voice interaction: Combined with iFlytek voice technology, it performs well in real-time translation and voice assistant scenarios.
• Education field: Applied to intelligent teaching and personalized learning plan generation.

Meta Model

Development organization: Tencent
Features:

• Multi-scenario coverage: Embedded in social platforms such as WeChat and QQ, supporting content generation, intelligent customer service and other functions.
• Synthetic data application: Reduce annotation costs and accelerate model iteration by generating synthetic data.

Pangu Model

Development organization: Huawei
Features:

• Industrial-grade applications: Focus on smart manufacturing, energy and other fields, and support the prediction and optimization of complex industrial scenarios.
• Full-stack technology: Combine Huawei Ascend chips and AI framework to provide end-to-end solutions.

GLM

Model name: GLM-4-Plus, GLM-4-9B
Development organization: Zhipu AI
Features:

• Efficient training: Reduce computing power requirements through distributed training technology, suitable for deployment in small and medium-sized enterprises.
• Academic cooperation: Cooperate with universities to promote open source ecology and support model customization in scientific research scenarios.

360

Model name: 360zhinao2-o1
Development agency: 360 Company
Features:

• Security-oriented: Focuses on threat detection and defense in the field of network security, and supports real-time data analysis.
• Low-cost API: Provides cost-effective API services, suitable for integration by small and medium-sized enterprises.

Doubao

Model name: Doubao-pro-32k-241215
Development agency: ByteDance
Features:

• Short videos and recommendations: Optimize video content understanding and recommendation algorithms to enhance the personalized experience of platforms such as Douyin.
• Multi-language support: Supports multi-language scenarios such as Southeast Asia and Europe to help international business.

Kimi AI

Model name: moonshot-v1-vision-preview, k1.5
Development organization: Darkside of the Moon Technology Co., Ltd.
Features:

• Ultra-long text processing: Supports 200,000 Chinese characters input, and the long text processing capability is 10 times the international top level, with outstanding performance in academic paper analysis, legal document analysis, API document understanding and other scenarios.
• Tool integration and cost optimization: Supports API Tool Calling function, and context cache technology reduces 90% of the long text processing cost.

Overseas large models

OpenAI

Model name: GPT-3.5 Turbo, GPT-4, GPT-4o, etc.
Development organization: OpenAI (USA)
Features:

• Multimodal capability: From GPT-3.5 to GPT-4, it gradually supports multimodal input and generation such as text, image, and voice.
• Industry benchmark: GPT-4 performs well in complex reasoning, mathematics, and programming tasks, with a total score of 80.4 (SuperCLUE list), especially leading the world in science tasks (87.3 points).
• Widely used: Integrated in products such as ChatGPT and Microsoft Copilot, covering content generation, code writing, educational assistance and other fields.

Claude

Model name: Claude 3.5 Haiku, Claude 3.5 Sonnet
Development organization: Anthropic (USA)
Features:

• Safety and controllability: Adopting the concept of "Constitutional AI", it reduces the generation of harmful content through preset rules and enhances user control.
• Complex reasoning advantage: Claude 3.5 Sonnet scored 54.6 in Hard tasks such as mathematics and code, which is close to the OpenAI model.
• Long context support: Supports ultra-long text input, suitable for scenarios such as legal document analysis and academic research.

Gemini

Model name: Gemini 2.0 Flash
Development organization: Google DeepMind (USA)
Features:

• Multimodal fusion: Supports full-modal processing of text, images, audio, and video, and performs outstandingly in cross-modal reasoning.
• Multi-language adaptation: Covers major global languages ​​and optimizes international scenario applications, such as real-time translation and multi-language content generation.
• Model stratification: Provides three versions: Ultra (super large model), Pro (general model), and Nano (lightweight) to adapt to different device requirements.

LLaMA（Meta AI）

Model name: LLaMA-3.3
Development organization: Meta (USA)
Features:

• Open source ecosystem: LLaMA-3.3-70B and other models are open source, attracting global developers to participate in improvements, and are widely used in intelligent customer service and text generation.
• Technical innovation: Uses rotational position encoding (RoPE) and SwiGLU activation functions to improve model performance and training stability.
• Vertical field optimization: Excellent performance in scenarios such as finance and healthcare, and supports localized deployment.

Mistral

Model name: Codestral, Mistral, Pixtral
Development agency: Hugging Face (France)
Features:

• Generative task optimization: Focuses on text and image generation, and is good at creative content production (such as advertising copy, art design).
• Lightweight deployment: The model parameters are moderately scaled, suitable for fast integration by small and medium-sized enterprises and developers.

Grok

Model name: grok-2-1212, grok-2-vision-1212
Development organization: Grok (USA)
Features:

• Industry deep adaptation: Outstanding performance in financial risk control and medical diagnosis, supporting high-frequency data analysis and pattern recognition.
• Real-time interaction: Optimize low-latency response, suitable for real-time decision-making scenarios (such as stock trading, emergency diagnosis).

Stable Diffusion

Model name: Stable Diffusion 3.5
Developer: Stability AI (UK)
Features:

• Image generation benchmark: Based on the diffusion model, it generates high-quality images and videos, which are widely used in artistic creation and film and television production.
• Open source community driven: Continuously iterate the model through community collaboration, supporting user-defined training and fine-tuning.

Synthesia

Development Agency: Synthesia (UK)
Features:

• Virtual Digital Human: Generates realistic virtual image videos, supports multi-language dubbing, and is used in education, advertising, and virtual customer service.
• Code-free operation: Users can quickly generate video content through text input, lowering the production threshold.

ElevenLabs

Model name: Multilingual v2, Flash v2.5
Development organization: ElevenLabs (USA)
Features:

• Voice cloning and synthesis: Clone the user's voice through a 15-second audio sample and support multi-language emotional voice generation.
• Cross-scenario application: Integrated in audiobook production, virtual assistants, game dubbing and other fields.