Powering the Agents: Workers AI Now Runs Large Models, Starting with Kimi K2.5

Why It Matters

By delivering frontier‑scale reasoning at open‑source pricing, Cloudflare removes the primary cost barrier to scaling enterprise and personal AI agents, accelerating broader AI adoption.

Key Takeaways

• Workers AI adds Kimi K2.5 with 256k context.
• Model cuts inference costs by ~77% versus proprietary LLMs.
• Prefix caching and session affinity boost throughput, lower token fees.
• New async API prevents capacity errors for batch agent workloads.
• Custom kernels deliver high GPU utilization without ML engineering.

Pulse Analysis

The AI landscape is rapidly shifting from proprietary giants to open‑source frontier models that rival commercial performance. Cloudflare’s integration of Moonshot AI’s Kimi K2.5 into Workers AI reflects this trend, giving developers access to a 256k context window and advanced tool‑calling capabilities without the hefty licensing fees. By embedding the model directly into its serverless edge platform, Cloudflare eliminates the need for separate hosting infrastructure, allowing teams to prototype, test, and deploy agents entirely within a unified environment.

Cost efficiency is a decisive factor for enterprises scaling AI workloads. Cloudflare reports that a security‑review agent processing 7 billion tokens per day saved roughly $2.4 million annually by switching to Kimi K2.5, a 77% reduction versus a mid‑tier proprietary model. This dramatic saving demonstrates how open‑source models can deliver comparable quality while dramatically lowering operational expenses, making continuous, high‑volume agentic tasks—such as code scanning, personal assistants, and real‑time analytics—economically viable at scale.

Beyond the model itself, Cloudflare introduced platform enhancements that address the practical challenges of serverless inference. Prefix caching with visible token metrics and a session‑affinity header reduces pre‑fill latency and improves token‑per‑second throughput, while the redesigned asynchronous API mitigates capacity bottlenecks for batch or non‑real‑time workloads. Custom inference kernels and advanced parallelization techniques further boost GPU utilization, delivering enterprise‑grade performance without requiring deep ML‑engineering expertise. Together, these innovations position Workers AI as a compelling choice for organizations seeking to deploy robust, cost‑effective AI agents across the edge.