Together AI optimizes MiniMax-M3 serving, boosting agentic workload throughput by up to 125% using custom sparse attention kernels · Digg