Cursor Reveals First Self-Trained Model Ships in Weeks, Launches Mobile iOS Beta

At a company event on June 23, Cursor co-founder Michael confirmed that the startup's first fully in-house trained AI model is in training and expected to ship within weeks. The Decoder reported that the model is built entirely from scratch — no open-source base like earlier Composer releases. In scale it is on par with Opus and GPT, and it uses between ten and twenty times more compute than any previous Cursor model. Notably, the model is designed to work beyond coding, extending Cursor's AI ambitions outside their core developer use case.

The collaboration with SpaceX on training was already known. Following SpaceX's $60 billion all-stock acquisition of Anysphere — Cursor's parent company — the two organizations confirmed they were jointly training a model on SpaceX's xAI Colossus supercomputing cluster. What is new today is the concrete timeline: training is active, the window is weeks, not quarters.

Cursor Mobile: Agent Supervision from iOS

Cursor also launched Cursor Mobile into iOS beta today. The app is not an attempt to replicate the full IDE on a phone. It is a supervision and control surface for agents running elsewhere — specifically, it lets users manage agents remotely, unblock stuck tasks, and review or comment on agent-generated screenshots. A remote control feature also provides access to agents running locally on a developer's machine.

This is a practical wedge for long-running agent workflows. As Cursor's agent mode takes on increasingly complex multi-hour tasks — refactoring across large codebases, running test suites, or autonomously filing PRs — the practical problem of "my agent needs input while I am away from my desk" becomes real. Cursor Mobile addresses it directly.

The move mirrors AWS's Kiro mobile coding supervision feature, announced the same week, suggesting the industry is coalescing around mobile-first agent management as a distinct category rather than a niche feature request.

Origin Update

The event also provided additional detail on Origin, Cursor's Git platform for both humans and AI agents, first announced at the Compile conference on June 19. Co-founder Tomas, who joined through the Graphite acquisition, confirmed Origin is running internally and with select partners today, with broad availability targeted for fall 2026. Origin handles merge conflicts, CI test failures, and PR comments natively, and is built on a new Git architecture running on top of cloud storage providers. In internal load tests, the team simulated thousands of agents reading from and writing to a single repository simultaneously.

What's Unconfirmed

Cursor has not committed to a specific release date beyond "weeks." Benchmark comparisons, pricing, and whether the model will be accessible through third-party API routers or remain Cursor-exclusive have not been announced. How the self-trained model will interact with Cursor's existing multi-model support — which currently includes Anthropic, OpenAI, Gemini, and xAI models — is also unclear. Cursor has not stated whether the self-trained model will replace or supplement the current model roster.