MiniMax-M2.1

MiniMax-M2.1 is a large-scale language model developed by MiniMax, designed for agentic workflows, multi-language programming, and complex real-world tasks. The model utilizes a sparse Mixture-of-Experts (MoE) architecture with approximately 230 billion total parameters and 10 billion active parameters per inference. This structure is intended to optimize inference speed and cost-efficiency while providing performance comparable to larger dense models.

A core focus of MiniMax-M2.1 is its comprehensive programming capability, extending beyond Python to include systematic optimization for languages such as Rust, Java, Golang, C++, Kotlin, Objective-C, and TypeScript. It is specifically engineered to handle the iterative cycles required for autonomous coding agents. Additionally, the model is tuned for what the creator calls "vibe coding," which emphasizes design comprehension and aesthetic precision in mobile and web development for platforms like Android and iOS.

The model introduces interleaved thinking, a feature that enhances its ability to follow complex instruction constraints and manage multi-step tool-calling processes. It has been evaluated using the VIBE (Visual & Interactive Benchmark for Execution) benchmark, which tests end-to-end application generation in live environments. MiniMax-M2.1 is released with open weights, supporting a native context window of up to 204,800 tokens.