Composer Is Free on Cursor (apidog.com)

Cursor has made its Composer coding agent widely available inside the Cursor IDE, pitching a fast, RL-trained mixture-of-experts (MoE) model that the team says matches frontier benchmarks while producing outputs roughly four times faster than comparable systems. Architecturally it uses sparse expert activation to cut compute, supports long-context generation for large codebases, and is trained with reinforcement learning to prefer verifiable, tool-driven actions. The stack includes PyTorch and Ray for asynchronous scaling across thousands of NVIDIA GPUs, expert parallelism and hybrid sharded data parallelism to minimize communication, and native low‑precision MXFP8 kernels so inference can run without post-training quantization. In practice the model runs as an agent that calls semantic search, grep and sandboxed terminals dynamically, integrates with IDE workflows and GitHub PRs, and can be paired with tools like Apidog for API generation and testing. For developers and the AI/ML community this matters because Composer demonstrates how MoE + RL + tool integration can deliver interactive, high‑quality code assistance at scale—improving developer flow, enabling safe command execution in sandboxes, and raising code quality via unit-test and linter feedback loops. The announcement also highlights accessibility: Cursor offers an official free tier, while the article notes community workarounds that claim to unlock paid features—approaches that may breach terms of service and carry legal/ethical risks. Technically, Composer is a useful case study in scaling sparse models and low‑bit training for real‑world software engineering agents.