小米于2026年9月21日发布MiMo-v2.6-Pro大语言模型1。该模型在Artificial Analysis Intelligence Index上得分46分，远高于该指标中位数18分1，展现出超越同类开源模型的性能表现。
MiMo-v2.6-Pro支持文本、图像、语音和视频等多模态输入1，具有1.0万亿总参数规模，其中活跃参数为42亿，采用混合专家（MoE）架构1。模型搭载1M tokens的上下文窗口1，推理速度达124.5 tokens/秒1，首token延迟为2.34秒1。定价方面，该模型的输入token成本为$0.43/百万，输出token成本为$0.87/百万1，均低于行业中位数的$0.45和$1.681。
模型采用链式思维推理技术1，以MIT许可证开源权重发布1，支持商业使用1。
Xiaomi has released MiMo-v2.6-Pro, a multimodal large language model that achieves strong performance metrics relative to comparable open-source alternatives 1. The model scored 46 points on the Artificial Analysis Intelligence Index, significantly exceeding the median score of 18 1. Released on September 21, 2026, the model supports text, image, speech, and video inputs while generating text outputs 1.
The MiMo-v2.6-Pro features a context window of 1.0 million tokens and delivers inference speeds of 124.5 tokens per second, outpacing the industry median of 74.8 tokens per second 1. Time to first token is 2.34 seconds 1. Constructed using a mixture-of-experts architecture, the model contains 1.0 trillion total parameters with 4.3 billion active parameters 1. The model incorporates chain-of-thought reasoning capabilities 1.
Pricing is set at $0.43 per million input tokens and $0.87 per million output tokens, both substantially lower than the median rates of $0.45 and $1.68 respectively 1. Xiaomi has open-sourced the model weights under an MIT license, permitting commercial use 1. The evaluation cost for benchmarking the model was $206.66 1.
评论
还没有评论，欢迎留下第一条。