MiniMax releases open-source M3.1 flash model with 428b parameters

On October 23, 2026, MiniMax announced the open-sourcing of its M3.1 Flash model, featuring 428 billion parameters and 23 billion active parameters. The model, launched on October 19, 2026, processed over 100 trillion tokens within 72 hours of its release. MiniMax highlighted significant improvements in model capabilities compared to its predecessor, M3, along with optimizations such as full model sparse attention, quantization-aware training, and power-aware scheduling to enhance inference efficiency. The company also partnered with major chip manufacturers and cloud platforms to support the open-source release. Hugging Face and ModelScope are among the platforms hosting the model.