Keep pulling the thread on Sualeh Asif.
Cursor's code indexing infrastructure is designed to operate at a scale of billions of files per day.
Cursor has been running the DeepSeek V2 model on its own inference infrastructure for approximately 8 to 12 months, scaling it to handle hundreds of millions of daily calls.
Cursor's custom "tab" model for code completion handles approximately 100 million requests per day, and this volume is growing quickly.
Cursor plans to evolve its infrastructure to privately store some user code in the future, which would enable faster and more reliable AI-driven edits.
Cursor uses Turbopuffer for its vector database, which stores vectors on Amazon S3.
The majority of Cursor's GPU compute cost is dominated by running its "tab" models for code completion, with indexing representing a very small percentage of total usage.
Cursor has been a long-time user of DeepSeek models, having previously used the 1.5 series and later switching to the V2 series.
Swaleh Asif considers DeepSeek models to be exceptionally well-pre-trained base models for building custom applications.
Anthropic's Sonnet model is particularly effective for use in AI agents because it maintains high coherence over long sequences involving tens of tool calls.
Cursor's mission is to automate as much of the coding process as possible while ensuring the developer remains in control.
GitHub Copilot did not improve significantly in the one to two years following its initial release.
The release of GPT-4 significantly raised the Cursor team's perception of the potential for building advanced AI coding products.