MiniMax M2.7 highspeed

The MiniMax M2.7 highspeed is built on a sparse Mixture-of-Experts architecture that utilizes 230 billion total parameters while activating only 10 billion per inference. This design intent focuses on achieving high-tier performance comparable to industry-leading models while maintaining significantly lower operational costs. By optimizing the active parameter count, the model provides a streamlined experience for demanding applications, particularly in software engineering where it has demonstrated strong results on technical benchmarks.

A defining characteristic of this model is its recursive self-evolution, where it autonomously performs over 100 iterations to refine its own training process, resulting in measurable performance gains without human intervention. This lineage of self-optimization allows the highspeed variant to achieve output rates of 100 tokens per second, representing a substantial increase in throughput. These advancements make it a practical choice for enterprise-scale deployments that require both rapid response times and high-level reasoning capabilities for complex development workflows.