Cursor publishes Composer 2 technical report, formally crediting Kimi K2.5 as base model

Cursor released a technical report on March 24 detailing how Composer 2 was trained, formally crediting Moonshot AI's Kimi K2.5 as the base model for the first time in an official document.^[1][2]

The report confirms what came to light days earlier when a developer discovered Kimi's model identifier in Cursor's API responses, prompting the company to acknowledge the open-source foundation it had omitted from its original launch announcement.^[3]

What the report reveals

The 23-page paper, authored by the Cursor Research Team, describes a two-phase training process: continued pretraining on code-dominated data followed by large-scale reinforcement learning.^[1]

In section 3, the report states that Cursor evaluated several open-source base models, including GLM-5, Kimi K2.5, and DeepSeek V3.2, before selecting Kimi K2.5, a 1.04 trillion parameter mixture-of-experts model with 32 billion active parameters.^[1] The selection was based on internal benchmarks measuring coding knowledge, state tracking, and codebase perplexity rather than agent-level coding benchmarks, which the team found less predictive of final performance.

Cursor then extended the model with continued pretraining on a large code dataset, followed by supervised fine-tuning and reinforcement learning using tasks drawn from real Cursor coding sessions. Training was performed on NVIDIA B300 GPUs.^[1]

The report also confirms Cursor's partnership with Fireworks AI for RL inference, consistent with what Lee Robinson and the official Kimi account stated on X after the initial disclosure.^[4][5] In the X thread announcing the report, Cursor thanked "the companies and open-source communities behind Kimi K2.5, Ray, ThunderKittens, PyTorch, and more" alongside Fireworks and Colfax.^[2]

Context

The report's explicit attribution of Kimi K2.5 contrasts with the original Composer 2 blog post on March 19, which described the model as delivering "frontier-level coding intelligence" without naming its base.^[6] That omission was flagged within hours by developer Fynn (@fynnso), who found the identifier kimi-k2p5-rl-0317-s515-fast in Cursor's API responses.^[7] Cursor co-founder Aman Sanger subsequently called the omission "a miss".^[3]

The technical report appears to be the corrective step the company signaled it would take. It names Kimi K2.5 in the introduction, provides detailed base model evaluation results in an appendix, and cites Moonshot AI's Kimi K2.5 paper in its references.