Humanoid robot Figure02 will be tested alpha in home environment in 2025
823
English
English
Instead of using a single AI model in a fixed manner, Cursor provides a flexible platform that supports multiple cutting-edge large language models (LLMs) to meet the needs of different developers. These models are integrated and optimized by Cursor’s team to provide efficient code generation, editing, and understanding. Here are some of the main models currently supported by Cursor.
1. Default model
Claude 3.5 Sonnet (provided by Anthropic): This is one of Cursor's default models, especially in recent versions (such as 0.39.1 update) that are set as the default choice for new users. It is known for its excellent code understanding and context processing capabilities, and is suitable for front-end development, complex project context reasoning and other tasks.
cursor-small : This is a lightweight model developed by the Cursor team, designed for fast completion and simple tasks. It is fast and has no usage restrictions and is suitable for small-scale editing in daily coding.
2. Premium Models
GPT-4 and GPT-4o (provided by OpenAI): These are powerful models, GPT-4o is the latest multimodal model of OpenAI, providing higher intelligence and performance for complex programming tasks and deep code analysis .
Claude 3 Opus and Claude 3.7 Sonnet (provided by Anthropic): Opus is an Anthropic high-end model with stronger inference capabilities, while 3.7 Sonnet is the latest version (supported by X post), allegedly in code generation and multiple Excellent in file editing.
Fusion : This is the next generation Tab model released by Cursor in January 2025 (see official website update), which is specifically used for predictive code editing and cursor jumping. It is based on a custom sparse language model, and after dozens of updates, it is faster and more accurate in suggestions.
3. User-defined model
Cursor allows users to add their own API keys through Settings > Models to access other models, such as OpenRouter or locally deployed open source models (such as Qwen2.5-Coder). This means you can use any supported external models as needed.
4. Model selection and switching
In the Cursor interface, you can switch models in Chat, Composer, Tab, or Agent modes through the drop-down menu. For example:
Chat and Composer : Use a 40,000 token context window by default, supporting long conversations.
Cmd-K : Limited to approximately 10,000 tokens to balance speed and mass.
Agent : Supports contexts up to 120,000 tokens, suitable for complex multi-file tasks.
Users can choose the most suitable model according to task requirements, such as Claude 3.5 Sonnet for front-end development and GPT-4o for complex reasoning.
5. What's special
Cursor not only relies on third-party models, but also conducts customized training on it. For example, the Tab function (formerly known as Copilot++) uses Cursor's own model, trained by billions of tokens, focusing on predicting code editing and cursor position.
It also provides functions such as .cursorrules files, allowing users to define project-specific rules and further optimize model output.
in conclusion
The AI model Cursor uses depends on your specific settings and subscription plan. Free users can access basic models (such as cursor-small and limited Claude 3.5 Sonnet), while Pro or Business users can unlock more advanced models (such as GPT-4o, Claude 3.7 Sonnet, and Fusion). If you want to confirm the latest models supported by the current version, you can check the Models page or update log of the Cursor official website.
