tl;dr: Cloudflare's Workers AI serves large MoE models like Kimi K2.6 and GLM 5.2 more efficiently by quantizing the KV cache to FP8 (doubling context capacity to ~1.37M tokens and boosting peak throughput ~41%) and compressing GLM weights to INT4 (shrinking the checkpoint 40% and speeding decode 16-55%), with negligible accuracy loss. They apply these selectively across disaggregated prefill/decode pools, and added a KV cache integrity check that catches page-mapping errors at under 1% overhead. All work is done on SGLang with patches upstreamed.