So Space Bunny was MiniMax M3.1 Flash all along! Honestly, the number that exc… · @GenAI_is_real · TwKit

@GenAI_is_real

So Space Bunny was MiniMax M3.1 Flash all along! Honestly, the number that excites me most isn't even the 74% on DeepSWE. It's the pricing: $0.15/M input tokens, $0.60/M output tokens, and just $0.015/M cached input tokens. Coding agents aren't one-shot benchmarks. They repeatedly read context, call tools, make mistakes, inspect failures, and try again. Making those iterations faster and cheaper can fundamentally change how much autonomy developers can afford. MiniMax also reports 3x the inference speed at roughly one-third of M3's agent-task cost. That's an exciting combination. Of course, the metric I'd really like to see is cost per successfully completed task, including retries, verification, and wall-clock time. Cheap tokens are nice. Cheap, reliable outcomes are much more interesting. And with open weights arriving on October 23, I'm especially excited to see what the inference community can do with it. Huge congrats to @MiniMax_AI and the team!

曝光 395 · 评论 3 · 点赞 3 · 书签 1