Chinese AI labs develop similar high-performance models

Two Chinese AI labs, Z.ai and Alibaba's Qwen team, have independently developed high-performance open-weight models with similar architectures. Z.ai's GLM-5.3-Flash, a 320B-parameter multimodal MoE model with 18B active parameters, was released alongside Qwen's Qwen3.8-Flash-Next, a 125B model with 6B active parameters. Both models use a 3:1 hybrid of linear and full attention, compressed indexers capped at 2048 tokens, and four gated residual branches. Z.ai claims GLM-5.3-Flash outperforms previous versions and rivals Claude Opus 4.8 in coding benchmarks. Qwen reports that training Qwen3.8-Flash-Next required about one-ninth the compute of its predecessor. Both models use the Muon optimizer and share similar design elements, though they differ in rotary position embeddings. MiniMax remains a notable dissenter, using full softmax attention in its models. The models' shared design elements suggest cross-pollination among Chinese open models, with DeepSeek and Moonshot AI contributing key components. Z.ai's GLM-5.3-Flash uses Kimi Delta Attention, while Qwen employs Gated DeltaNet. Both models compress context 4x before scoring and cap attention at 2048 tokens. Qwen credits its QSA design with significant speedups. Despite similarities, MiniMax's M3 model uses a different approach, highlighting ongoing debates in the field. Sources include Hugging Face, GitHub, and academic papers. The models' release marks a significant step in the evolution of large-scale AI architectures, with implications for efficiency and performance.