Cloudflare's Workers AI now supports large frontier-scale models, launching with Moonshot AI's Kimi K2.5 (256k context, multi-turn tool calling, vision inputs). Cloudflare tested the model internally for agentic coding tasks and automated code review, reporting a 77% cost reduction versus mid-tier proprietary models for a security review agent processing 7B tokens/day. To support large models, Cloudflare built custom kernels on their Infire inference engine and implemented techniques like disaggregated prefill and tensor/expert parallelism. Two new platform features accompany the launch: prefix caching with session affinity headers (via x-session-affinity) to reduce TTFT and costs on multi-turn conversations, and a revamped asynchronous API using a pull-based queue system to handle high-volume, non-real-time workloads without capacity errors.
