Side effect of blocking Chinese firms from buying the best NVIDIA cards: top models are now explicitly being trained to work well on older/cheaper GPUs. The new SoTA model from @Kimi_Moonshot uses plain old BF16 ops (after dequant from INT4); no need for expensive FP4 support.

•

Zhihu Frontier

@zhihufrontier · Twitter · 2025-11-08 11:50 UTC

🚀 "Quantization is not a compromise — it's the next paradigm." After K2-Thinking's release, many developers have been curious about its native INT4 quantization format. 刘少伟, infra engineer at @Kimi_Moonshot and Zhihu contributor, shares an insider's view on why this choice