MiniMax要开源2.7T参数的巨模型，比M3大六倍多，专攻复杂推理。虽然本地跑不动，但开源权重让研究者能动手玩，看看中国厂商怎么卷。
MiniMax计划在Q3开源代号M3 Pro的模型，参数量达2.7T，是当前旗舰M3（428B参数）的6倍多。新模型在复杂推理和多步任务上有显著提升。由于规模巨大，M3 Pro仅适合在专业基础设施上运行，无法在消费级硬件上实时使用。开源权重可供社区实验、微调或蒸馏更小模型。
MiniMax is planning a 2.7T parameter open source m…
MiniMax is planning a 2.7T parameter open source model.
Their current flagship M3 sits at 428B parameters. The new one, codenamed M3 Pro, is expected to drop open source in Q3 with gains on complex reasoning and multi-step work.
Serving a model this size will mostly happen through providers running it on proper infrastructure — local running on consumer hardware looks unrealistic for anything close to real-time use, even with heavy quantisation and tricks like memory mapping.
The open weights are still valuable though: they let others experiment, fine-tune, or pull out better smaller models. It’ll be worth watching how quickly other labs, especially in China, answer with their own large releases or efficiency improvements. https://t.co/OeEgSAnU4v