Cloudflare has adopted Kimi 2.5 as the foundational model for its internal engineers' code agent. Cloudflare (a publicly traded company with a market cap of around $77 billion) initially started experimenting by using Kimi K2.5 as an internal coding agent. Engineers use it daily to execute tasks in the OpenCode environment and have integrated it into the automated code review process, monitoring its performance through GitHub's Bonk agent. It processes 7 billion tokens daily. Previously, using large proprietary models (like OpenAI or Anthropic) cost $2.4 million annually just to maintain a single codebase. Switching to Kimi K2.5 + Workers AI instantly reduced costs by 77%. A few days ago, Cursor's Composer 2, which is essentially a wrapper around Kimi, also started from Kimi K2.5. Through RL training, its coding capabilities are now sufficient for everyday development. Looks like Kimi 2.5 is proving to be quite capable in coding scenarios. Cursor really commercialized the open-source community's work (winning big). Composer 2's base model uses Kimi 2.5, and the RL training seems mainly focused on compressing long-range context. But looking at their technical blog: https://(cursor.com)/blog/self-summarization, it’s clear there’s not much real innovation of their own. The open-source community has long had many RL-based context compression papers and solutions, and it seems they’ve referenced a lot of them. https://blog.(cloudflare.com)/workers-ai-large-models/