Powering the agents: Workers AI now runs large models, starting with Kimi K2.5 (blog.cloudflare.com)

Cloudflare has officially entered the large model arena with its launch of the Kimi K2.5 model on the Workers AI platform, enhancing its capabilities for deploying AI agents. Kimi K2.5 supports a significant context window of 256k tokens and incorporates features for multi-turn tool calling and structured outputs, making it well-suited for a variety of complex tasks. This new offering allows developers to manage the entire lifecycle of AI agents within a single platform, leveraging existing Cloudflare primitives such as Durable Objects and Workflows for a coordinated infrastructure. The significance of Kimi K2.5 in the AI and ML community lies in its cost efficiency and performance capabilities. Cloudflare reports a dramatic 77% reduction in costs for their internal security review agent, which processes over 7 billion tokens daily, when transitioning from proprietary models to Kimi K2.5. The platform also boasts improved performance techniques such as custom kernels and prefix caching, which enhance GPU utilization and throughput. With a focus on affordable, open-source AI solutions, Cloudflare aims to empower enterprises and individuals alike, facilitating a shift towards accessible AI agent deployment without the steep costs of proprietary frameworks.