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

Cloudflare's Workers AI now supports large frontier open-source models, launching with Moonshot AI's Kimi K2.5 to power agentic workloads on a unified platform.

• •Kimi K2.5 offers a 256k context window, multi-turn tool calling, vision inputs, and structured outputs, making it suitable for agentic tasks.
• •Internal usage showed a 77% cost reduction compared to mid-tier proprietary models; a security review agent processing 7B tokens/day would cost $2.4M/year with proprietary models vs. a fraction with Kimi.
• •Workers AI built custom kernels on its proprietary Infire inference engine, using tensor/expert parallelization and disaggregated prefill to optimize large model serving.
• •Prefix caching is now surfaced as a usage metric with discounted pricing; a new x-session-affinity header improves cache hit rates, reducing TTFT and inference costs.