Cursor, the vibe‑coding platform from Anysphere, has rolled out Composer as part of its latest Cursor 2.0 update. Composer is the company’s first proprietary coding LLM and promises a four‑fold speed improvement over similarly intelligent competitors while preserving robust reasoning over large, complex codebases. The model finishes most requests in under 30 seconds, matching the performance of fast‑inference models and outpacing front‑line systems in token‑generation speed (≈ 250 tokens per second). The boost comes without sacrificing accuracy: Composer’s internal “Cursor Bench” suite—built from real developer agent requests—shows it aligns with existing abstractions, style conventions, and engineering best practices.
What sets Composer apart is its training regimen. Lead researcher Sasha Rush explains that the team used reinforcement learning to fine‑tune a mixture‑of‑experts (MoE) architecture. Rather than feeding the model static code or text, Cursor immersed Composer in full codebases, complete with file editing, semantic search, terminal commands, and unit‑test execution. Each training iteration required the model to plan, write, test, and review code, encouraging emergent behaviors like linter fixes and multi‑step searches. The RL loop optimized for both correctness and efficiency, teaching Composer to choose tools judiciously, parallelize tasks, and avoid speculative outputs. This real‑world, production‑scale training means Composer can operate directly within a developer’s IDE, handling version control, dependency resolution, and iterative testing just as a human engineer would.
Composer is tightly integrated into Cursor 2.0’s agentic environment. The platform now supports up to eight isolated agents, each running in its own workspace via git worktrees or remote machines. Composer can function as a single agent or participate in collaborative runs, allowing developers to compare multiple outputs side‑by‑side. Additional features—such as an in‑editor browser, sandboxed terminals, improved code review, and voice mode—further streamline the workflow and keep the model’s response time snappy. Enterprise customers receive administrative controls, audit logs, and SAML/OIDC authentication, while pricing tiers range from free hobby plans to $200/month Ultra tiers for heavy users.
In a landscape where many AI assistants act as passive suggestion engines, Composer represents a shift toward continuous, autonomous coding. By training inside a live coding ecosystem and coupling that training with an RL‑driven MoE design, Cursor delivers a model that is not only fast but also context‑aware and production‑ready—an early glimpse of what everyday software development could look like when humans and AI share the same workspace.
Want the full story?
Read on VentureBeat →