MiniMax Unveils M1 AI Reasoning Model, Slashing Compute Use vs DeepSeek-R1
MiniMax 推出 M1 人工智能推理模型,相比 DeepSeek-R1 大幅削减计算用量
Shanghai’s MiniMax debuts open-source M1, its first reasoning model, claiming it uses under half the computing resources of DeepSeek-R1 for tasks with ≤64K tokens. Built on a 456B-parameter foundation, M1 employs hybrid MoE and Lightning Attention, offering 10x larger context window. Third-party tests show it matches top global models in math/coding, while MiniMax’s paper references DeepSeek 24 times, highlighting rivalry.
上海 MiniMax 发布开源 M1—— 其首款推理模型,称在处理≤6.4 万令牌任务时,计算资源消耗不到 DeepSeek-R1 的一半。M1 基于 4560 亿参数基础模型,采用混合专家系统和 Lightning Attention 技术,上下文窗口达 10 倍。第三方测试显示其在数学 / 编程领域性能媲美全球顶尖模型,而 MiniMax 论文 24 次提及 DeepSeek,凸显竞争态势。
