MiniMax's new open M2.5 and M2.5 Lightning near SOTA, costing 1/20th Opus 4.6 (venturebeat.com)

Chinese AI startup MiniMax has launched its M2.5 language model in two variants, promising revolutionary pricing that reduces costs of high-end AI to as low as 1/20th of its competitors like Claude Opus 4.6. Open-sourced on Hugging Face, M2.5 utilizes a Mixture of Experts (MoE) architecture, activating only 10 billion of its 230 billion parameters per task for enhanced efficiency, and was trained using a proprietary Reinforcement Learning framework called Forge. This model, designed with input from professionals across various fields, allows for significant cost reductions, with operational expenses around $10,000 annually for four AI workers, making sophisticated AI accessible for enterprises. The significance of M2.5 lies not just in affordability, but in its capacity to perform complex autonomous tasks, transitioning AI from simple chatbots to integrated "agents" capable of handling real-world work independently. With high benchmark scores rivaling industry leaders, MiniMax positions M2.5 as a transformative tool for enterprises, enabling high-reasoning processes without the anxiety of escalating costs. The model's development encourages innovation in how organizations utilize AI, promising efficiency and operational scalability while maintaining control over data privacy through its open-source nature.