March 20, 2026AgentsCodingOpen Source

Cursor Composer 2: Frontier-Level Coding Model Built on Kimi K2.5 with RL

Cursor released Composer 2 on March 19, 2026 — a new coding model that delivers frontier-level performance on challenging coding tasks. Priced at $0.50/input and $2.50/output tokens (with a faster variant at $1.50/$7.50), Composer 2 introduces "compaction-in-the-loop reinforcement learning" that pauses generation to compress context to roughly 1,000 tokens, enabling efficient long-context coding.

However, the launch quickly became controversial. Developers identified the model ID as "kimi-k2p5-rl-0317-s515-fast," revealing that Composer 2 is built on Moonshot AI's Kimi K2.5 with additional reinforcement learning training. Cursor's blog post highlighted RL-driven performance improvements but omitted any mention of Kimi K2.5 as the base model.

Moonshot AI's head of pretraining Yulun Du confirmed the tokenizer similarity and questioned Cursor's compliance with the model's modified MIT license, which requires commercial products to prominently display "Kimi K2.5" branding. Cursor, valued at $29.3 billion, has not responded to the allegations.

The story highlights a growing tension in the AI coding agent space: as companies build commercial products on open-weight models, attribution and licensing compliance become critical issues. Composer 2 is available now on Product Hunt and through the Cursor editor.

Official blog: https://cursor.com/blog/composer-2
HN discussion: https://news.ycombinator.com/item?id=47452404
← Previous
Microsoft Foundry IQ: Knowledge Grounding for AI Agents via MCP
Next →
Google AI Studio 2.0: Full-Stack Vibe Coding with Antigravity Agent and Firebase
← Back to all articles
>_