Powering the agents: Workers AI now runs large models, starting with Kimi K2.5

This post announces the integration of large open-source models, specifically Kimi K2.5, into Workers AI, enabling cost-effective and high-performance inference for agentic workloads. It details the underlying inference stack optimizations, including custom kernels and parallelization techniques. New platform features for agentic workloads include prefix caching with surfaced metrics and discounts, a new `x-session-affinity` header for improved cache hit rates, and a redesigned asynchronous API for durable inference. The post quantifies cost savings (77% reduction for a specific use case) and performance benefits (faster TTFT, higher TPS) derived from these advancements.