This technical comparison evaluates three open-weight language models released by Chinese AI labs in 2026. Kimi K3, at 2.8 trillion parameters, achieves the highest benchmark scores but remained API-only until its late-July weight release, while DeepSeek V4 Pro, at 1.6 trillion parameters, offers the lowest serving costs at approximately $0.04 per task. GLM-5.2, at 744 billion parameters, provides the fastest inference speed and is immediately self-hostable under an MIT license, positioning it as a middle ground between K3’s capability and V4 Pro’s cost efficiency.
