Chinese lab ZAI cuts costs 10x by adopting peer innovations like DeepSeek

PAstynome · x · 2026-08-26

The post argues that export controls may inadvertently incentivize Chinese AI labs to develop cost-reducing innovations. Citing ZAI (Zhipu) as an example, it highlights how quickly they adopted techniques from peers like DeepSeek and MoonShot, including mHC-based residual connections, DSA-like sparse attention, and linear attention. The result is a model with strong performance that uses 4.44x less KV cache and 3x fewer FLOPS, serving 10x cheaper than their previous model, demonstrating the economic efficiency of building on shared innovations.

Original post →

More from Infra

Infra channel →