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

A faster model only helps if the finished job gets faster. Morgan Suo of MiniMax uses one real research task to test the changes inside the M3.1 Flash model: an analyst checking whether circuit-board suppliers are turning AI demand into cash. The changes are sparse attention on long inputs, a 4-bit KV cache, low-precision experts, speculative decoding, server load and reasoning budgets. Each one saves compute, and each can quietly cost accuracy or wait time. She then shows how to judge a model switch: check sources and calculations, count the full cost of a usable report, and make the result r