Powering Agents: Workers AI Starts Supporting Large Models Beginning with Kimi K2.5

Cloudflare has begun rolling out frontier-class open-source large models on Workers AI in earnest. The first model is Moonshot AI's Kimi K2.5, which supports a 256k context window, multi-turn tool calling, vision inputs, and structured outputs, making it well-suited for agentic tasks.

Cloudflare has already been offering execution primitives such as Durable Objects, Workflows, Dynamic Workers, and Sandbox, along with the Agents SDK, but now models can be handled within the same platform as well. This means the entire lifecycle and execution environment of an agent can be operated together on a single development platform.

Cloudflare also emphasized the utility of Kimi K2.5 in terms of performance and cost. Internally, Cloudflare uses this model for everyday agentic coding tasks in its OpenCode environment, and has applied it to its automated code review pipeline and its public code review agent, Bonk. One security review agent processes over 7B tokens per day and has found 15+ confirmed issues in a single codebase; Cloudflare stated that switching to Kimi from a mid-tier proprietary model, which would have cost $2.4M annually, achieved a 77% cost reduction.

To reliably serve large models, Cloudflare also reworked its inference stack. It applied custom kernels for Kimi on top of the Infire inference engine to boost performance and GPU utilization, and integrated techniques used for optimizing large models—such as data parallelization, tensor parallelization, expert parallelization, and disaggregated prefill—tuned for production. Cloudflare explains that the platform handles the kind of optimization work that would otherwise require ML Engineer, DevOps, and SRE-level effort if self-hosting.

Platform features have also been strengthened with agentic workloads in mind.

• Leveraging prefix caching to reuse input tensors from previous requests, achieving faster TTFT and higher TPS.
• Exposing cached tokens as a usage metric, billed at a lower rate than input tokens.
• Introducing the x-session-affinity header to improve per-session cache hit rates.
• For non-real-time workloads, offering a newly revamped Asynchronous API that processes queued requests sequentially as capacity becomes available.

Cloudflare sees this asynchronous approach as particularly useful for workloads that don't require an immediate response, such as code-scanning agents or research agents. Alongside the model page, it has also opened up prompt caching, the batch API, an Agents SDK starter, OpenCode integration, and a playground, pushing Kimi K2.5 forward as the default agentic model on Workers AI.