On October 23, MiniMax officially open-sourced the M3.1 Flash model weights. The company also announced that the model processed over 1 trillion tokens within 72 hours of going live and will collaborate with several chip manufacturers, inference platforms, and cloud service providers to facilitate access on the first day of open-source release. According to MiniMax, M3.1 Flash was released on October 19, primarily targeting applications in agent tasks, programming, computer operations, and cybersecurity.
According to data released by MiniMax, the online inference speed of the model has improved nearly threefold compared to previous versions, with usage costs reduced by 70%. The model employs a sparse attention architecture, achieving a daily output speed of over 200 tokens per second, with a median output speed consistently above 170 tokens per second.
M3.1 Flash model significantly improves performance and reduces costs
In terms of API pricing, M3.1 Flash charges 1 yuan per million tokens for input, 4 yuan for output, and 0.1 yuan for cache reads. Compared to the corresponding prices of M3, both input and output unit prices have been reduced by over 50%, while the cache read unit price has decreased by more than 70%.
MiniMax stated that it will collaborate with chip manufacturers such as Huawei Ascend, T-head, Kunlun Core, Moore Threads, NVIDIA, and AMD, as well as relevant inference and cloud service platforms, to promote the integration of M3.1 Flash. By open-sourcing the model weights, developers can deploy and optimize according to their own business and hardware environments.
It was noted that the Space Bunny model, which previously operated anonymously and processed over 200 trillion tokens, is an early version of M3.1 Flash. MiniMax also announced that a Pro version with higher intelligence capabilities will be launched in the future, although a specific release date has not yet been disclosed.