MiniMax open-sources M3.1 Flash model with 428B parameters

Chinese AI company MiniMax has open-sourced the weights of its MiniMax M3.1 Flash model, with 428B total parameters and 23B activated parameters. Compared to the previous M3 model, the new version delivers improvements across the board. The model launched on October 19 and handled over 100 trillion tokens from real users within 72 hours. M3.1 Flash boosts inference efficiency through three key optimizations: replacing the first three full attention layers with sparse attention to cut computing costs, using quantization-aware training (QAT) to move MoE experts to W4A4 precision, and fine-tuning chip power allocation to keep computing cards running at high performance. MiniMax has partnered with chip makers like NVIDIA, AMD, and Huawei Ascend, plus cloud platforms such as Together AI and Fireworks to support the model on launch day. Open source communities like vLLM, SGLang, and Hugging Face are also helping bring the model to developers worldwide.