Agentic serving and model-internal computation raise systems questions

TAO jointly schedules LLM inference around tool readiness, cache value, and token footprint, reporting up to 26.8% lower latency and 29.5% higher decode throughput across 12 settings; deployment impact depends on workload and baseline validation. A separate report of a RISC-V CPU compiled into Qwen 3.5 weights is an intriguing model-computation result, but its claims need independent reproduction.