Minimax M3.1 Flash is 428B 23a

Explain Like I'm Normal

Chinese AI lab MiniMax quietly revealed that their new M3.1 Flash model boasts a massive 428 billion parameters, despite the 'Flash' naming usually implying a smaller footprint. It utilizes a Mixture-of-Experts architecture, activating about 23 billion parameters per token to balance speed with high-capacity reasoning. This makes it one of the largest open-ish weights models currently available, though running it locally requires massive VRAM.