Home / Technology / AI Gets Dirt Cheap: New Model Slashes Costs
AI Gets Dirt Cheap: New Model Slashes Costs
13 Feb
Summary
- New M2.5 model drastically cuts AI costs by up to 95%.
- AI is shifting from chatbot to affordable 'worker' roles.
- M2.5 uses MoE architecture, activating fewer parameters.
Chinese AI startup MiniMax, based in Shanghai, has unveiled its M2.5 language model, designed to significantly lower the cost of advanced AI. The model offers performance comparable to top-tier AI from Google and Anthropic but at a substantially reduced price, potentially cutting costs by as much as 95%.
This release signifies a move towards AI as a 'worker' rather than just a 'chatbot.' MiniMax has integrated M2.5 into its own operations, with it handling 30% of tasks and 80% of new code generation. The company trained the model using a proprietary Reinforcement Learning framework called Forge over two months.
The M2.5 model leverages a Mixture of Experts (MoE) architecture, activating only 10 billion out of its 230 billion parameters for each word generated. This approach allows for high reasoning capabilities with greater agility and efficiency.
MiniMax claims that M2.5 can handle specialized industry tasks, showing strong performance in financial modeling. The model's speed improvements also enable 'agentic' pipelines for real-time applications. Positioned as open-source, it promises scalable automated code audits and enhanced data privacy, pending official licensing details.
