Cloudflare Runs Kimi and GLM Models on Workers AI at Scale

Cloudflare has added Kimi (from Moonshot AI) and GLM (from Zhipu AI) to its Workers AI inference platform. The announcement focuses on the operational tradeoffs that matter when running these models at scale: model size, latency, and safety guardrails baked into the serving layer.

Both models are positioned as smaller and faster alternatives to frontier-class models. That framing is deliberate. Smaller parameter counts reduce memory pressure per inference request, which translates directly to lower cost and tighter latency on distributed edge infrastructure. For builders running high-throughput pipelines or latency-sensitive applications, that profile is often more useful than raw capability benchmarks.

GLM from Zhipu has strong multilingual coverage, particularly in Chinese, making it a practical fit for applications targeting East Asian markets without routing traffic through US-based inference endpoints. Kimi, developed by Moonshot AI, has built a reputation for long-context handling — a property that holds value in document processing, retrieval-augmented generation pipelines, and agentic workflows where context windows fill quickly.

The Cloudflare edge deployment means inference requests resolve close to the user, reducing round-trip latency without requiring teams to manage GPU infrastructure. Workers AI handles the serving, scaling, and model versioning. Developers call a REST endpoint or use the AI binding inside Workers scripts.

The "safer" framing in the announcement points to inference-layer controls rather than model fine-tuning. Cloudflare applies its own filtering and safety tooling at the serving layer, separating model behavior from platform-level guardrails. That separation matters for teams building on regulated or consumer-facing products.

For solo founders and small teams, access to Kimi and GLM through Workers AI removes the friction of standing up separate inference infrastructure for non-OpenAI models. The catalog now covers a broader range of capability-cost tradeoffs without requiring vendor sprawl.

Source