Ling-Plus

Released in March 2025 alongside Ling-Lite, Ling-Plus is a 290-billion-parameter Mixture-of-Experts model with 28.8 billion activated parameters, trained by Ant Group's Ling team on heterogeneous, lower-specification hardware rather than premium H100/H800 clusters. The team reported roughly 20% lower training compute costs versus high-end hardware, with performance comparable to leading models of similar scale on Chinese and English benchmarks.

Limitations: performance claims are vendor-reported and have not been independently verified at scale; the 290B MoE still requires substantial inference hardware despite the cheaper training story.