SOTAAZ: KV cache, measured — Qwen3.5-9B hybrid vs Qwen3-8B on one A100

Third-party measurement, read 2026-10-18. Rig: one A100 80GB PCIe, driver 535.288.01, CUDA 12.2, vLLM 0.28.0 (+cu129), transformers 5.16.1. Qwen3.5-9B (hybrid Gated DeltaNet, 32 layers, full attention every 4th) and Qwen3-8B (dense GQA) both in bf16 at max_model_len 40960, gpu_memory_utilization 0.90, batch 1. Reports Qwen3.5-9B decode 93.9 tok/s at 32K context and 69.7 tok/s at 40K, against Qwen3-8B 63.4 and 56.3 tok/s at the same points, and a 3.6x concurrency advantage (35.1 vs 9.77 concurrent 40K requests) attributed to the hybrid cache. Also a prefix-reuse sweep on Qwen3-8B: 1.00x at 512 shared tokens, 1.62x at 2K, 3.02x at 8K, 9.58x at 32K shared tokens, and 0.996x once the working set exceeds the cache. The post is a blog, but it publishes the rig, the software versions and the raw-JSON location, so it is recorded as a measurement rather than a marketing claim.