Is Cursor using their own in house model? What have experts said about it?

Cursor employs a sophisticated and pragmatic multi-model strategy rather than relying on a single in-house model [1, 29]. The company uses an ensemble of different large language models, with one expert citing as many as 13 distinct AIs, to power various aspects of its code editor [3, 17]. This hybrid approach involves leveraging large, powerful third-party foundation models from providers like Anthropic for complex reasoning and agentic tasks, while deploying a suite of custom-trained and fine-tuned models for performance-critical functions [1, 3, 15]. Initially, the team planned to rely on third-party models, but they found it essential to develop their own to achieve the necessary product quality and build a competitive moat, thereby challenging the notion that AI application companies are merely "wrappers" .

The company's in-house and custom model development is extensive and operates at a significant scale. Cursor runs the DeepSeek V2 model on its own inference infrastructure, which it has scaled to handle **hundreds of millions of daily calls** [4, 8, 14]. For code completion, a custom "tab" model processes approximately 100 million requests per day . A key proprietary asset is the "Composer" model, described as a fine-tuned version of a large Chinese mixture-of-experts model that is designed for rapid coding iterations [2, 28]. This model is particularly dynamic, with one source reporting that its weights are updated based on real-world user feedback as frequently as every 90 minutes . In addition to these specialized models, Cursor has also released its own foundation model specifically designed for programming, which serves as an alternative to models from OpenAI or Anthropic within the platform .

The strategic rationale for this custom model development centers on performance, defensibility, and unit economics. For speed-sensitive tasks like autocomplete, custom models are crucial for meeting latency targets of **under 300ms** . This focus is so integral that experts report every significant feature or "magic moment" in the editor is powered by a custom-trained model [23, 25]. This strategy also creates a data flywheel, where user data from larger models is used to distill smaller, faster, and more efficient versions . Economically, developing proprietary models is a key lever for improving margins. One analyst noted that Cursor's gross margins are low due to payments to model providers like Anthropic , and another expert described the strategy as achieving massive distribution with third-party models before launching an efficient in-house model like Composer to drastically improve profitability [9, 22]. This allows Cursor to build a more durable business by creating specialized models for specific coding tasks, particularly for enterprise users .