Cloudflare Workers AI adds Kimi K2.5 for larger agent workloads
Cloudflare says Workers AI now runs Kimi K2.5, giving builders a hosted open-source model with long context, tool calling, vision inputs, and structured outputs for agent workflows.
Cloudflare Workers AI adds Kimi K2.5 for larger agent workloads
Cloudflare says Workers AI now supports Kimi K2.5, a frontier-scale open-source model aimed at agentic workloads. The practical change is that teams can run a long-context, tool-calling, vision-capable model inside Cloudflare's developer stack instead of splitting agents across separate model, execution, and workflow vendors. Cloudflare's documentation lists the hosted model as @cf/moonshotai/kimi-k2.5 with a 256,000-token context window, function calling, reasoning, vision, batch support, and published unit pricing.
Key takeaways
- Workers AI now lists Kimi K2.5 as a hosted Moonshot AI text-generation model for agentic workloads.
- Cloudflare positions the release as part of a unified agent stack alongside Workers, Durable Objects, Workflows, Dynamic Workers, Sandbox, and the Agents SDK.
- The model documentation shows a 256k context window, tool/function calling, vision support, structured-output use cases, and pricing of $0.60 per million input tokens and $3.00 per million output tokens.
- Moonshot's own Kimi K2.5 post confirms the model is multimodal and agent-oriented, with Kimi.com, API, app, and Kimi Code availability.
- The most important limitation is operational: long context does not remove the need to control tool permissions, logging, cost caps, and output verification.
Practical LinkLoot angle
This is useful if you are building agents that need model inference, code execution, state, and workflow orchestration in one deployment path. A simple pattern is: run the user-facing agent on Workers, store durable task state in Durable Objects, put multi-step jobs in Workflows, use Sandbox or Dynamic Workers for risky execution, and call Kimi K2.5 through Workers AI when the task needs long context or vision input.
| Option | Best use | Limitation | Source |
|---|---|---|---|
| Workers AI + Kimi K2.5 | Hosted open-source model inside Cloudflare's agent platform | Cloudflare docs list a planned deprecation date, so production teams should monitor model lifecycle notices | Cloudflare docs |
| Kimi.com / Kimi API | Direct access to Moonshot's own Kimi modes and Kimi Code | Separate from Cloudflare's runtime, state, and deployment primitives | Moonshot AI |
| External proprietary model API | Highest-model-choice flexibility | More integration work for auth, tracing, data movement, and execution isolation | Vendor docs |
For LinkLoot readers, the immediate workflow is not "switch every agent to Kimi." It is to test one contained task: a support triage agent, repository summarizer, visual QA assistant, or internal research workflow where long context and tool calling reduce handoffs.
What to verify before you act
Check the model lifecycle first: Cloudflare's model page lists a planned deprecation date, so do not hard-code the model name without a fallback. Then test your real workload against the published pricing, because long-context prompts can still become expensive if logs, retrieval chunks, or tool outputs are appended without pruning. Finally, verify tool permissions and audit logs before letting an agent call browsers, shells, databases, or deployment APIs from a production Worker.
Source check
The Cloudflare announcement confirms that Workers AI is adding Kimi K2.5 and frames it as part of Cloudflare's agent infrastructure. Cloudflare's model documentation confirms the hosted model name, context size, function calling, vision, batch support, pricing, and model metadata. Moonshot AI's Kimi K2.5 technical blog independently corroborates that Kimi K2.5 is a multimodal, agent-oriented model and describes its Kimi.com, API, app, and Kimi Code availability.
It is Cloudflare's hosted Workers AI listing for Moonshot AI's Kimi K2.5 model, exposed as @cf/moonshotai/kimi-k2.5.
If you are comparing agent stacks, start with LinkLoot's guide to AI workflow automation and map each task to model, state, tools, approval gates, and rollback paths before choosing a platform.