AWS launched the Trainium3 chip in EC2 UltraServers, offering 4.4x performance for AI training and 40% better efficiency, supporting massive 1M-chip clusters.
MoE
Moonshot AI open-sourced its Kimi K2 model, a 1T-parameter MoE AI, showing exceptional performance (outperforming Grok) and prompting Perplexity AI to use it for post-training.