Prefill-first inference for open models

Starting with real-time voice

200ms TTFT on GLM-5.2

80ms TTFT on Qwen3.6-35B-A3B

BLOG