MiniMax-M3.1-Flash

MiniMax-M3.1-Flash is a native multimodal mixture-of-experts (MoE) language model developed by MiniMax. Initially trialed anonymously under the pseudonym "Space Bunny" on public leaderboards, the model was officially launched in October 2026. It is designed to excel in complex agentic workflows, software development, computer use, and cybersecurity tasks.

Architecture and Efficiency

The model features an architecture with approximately 428 billion total parameters, with 23 billion parameters activated per token. Building upon its predecessor MiniMax-M3, M3.1-Flash incorporates MiniMax Sparse Attention mechanisms throughout its network, quantization-aware training, and DSpark speculative decoding to optimize inference speed and minimize operational expenses. It supports text, image, and video inputs alongside an expansive context window of up to 1 million tokens.