Training the biggest AI models just got 1.8 times faster with less memory

A new method makes training the largest AI models much more efficient. Developers can now build these complex models faster, using less computing power and memory.

Building the biggest AI models, called Mixture of Experts, costs a lot of money and hardware. This new method cuts those costs significantly. It means companies can develop larger, more capable AI models without needing to buy proportionally more expensive hardware. This lowers the barrier for anyone trying to build state-of-the-art AI.