minimax-m3.1IMAGE
M3.1 Flash 官方跑分：Automation-Bench 62.8 超 Opus 5，Terminal Benc
作者列出 428B/23B 激活、1M 上下文与开源权重，指出官方表中 Automation-Bench 62.8 对 Opus 5 的 50.3，但 Terminal Bench 4.0 与 OSWorld 2.0 落后，并问哪个跑分更贴近真实工作。
ponster.sol
@ryanpon3 天前收录原推发布于 2026.10.25
ponster.sol @ryanpon原推发布于 2026.10.25
MiniMax M3.1 Flash: 428B params, 23B active per token, 1M-token context, weights on Hugging Face. Its own table has it beating Opus 5 on Automation-Bench (62.8 vs 50.3) but trailing on Terminal Bench 4.0 and OSWorld 2.0. Which benchmark best predicts your real work?
x.com/ryanpon/status/2114148426862543177采集于 2026.10.25 12:07