Is a Faster Model Actually Faster? Test the Whole Job — Morgan Suo, MiniMax

Why watch

Morgan Suo from MiniMax demonstrates how to properly evaluate whether a faster AI model actually delivers faster results using a real circuit-board supplier analysis task. She walks through specific optimizations in the M3.1 Flash model—including sparse attention on long contexts, 4-bit KV cache, low-precision MoE experts, speculative decoding, and reasoning budgets—explaining how each trades accuracy or wait time for compute savings. The video teaches AI engineers and ML practitioners how to measure the full cost of getting a usable report, not just token generation speed.