Z.ai's GLM-5.2 Introduces IndexShare for Efficient Long-Context Processing

Z.ai released GLM-5.2, an open-weight model that initial assessments indicate is the strongest open-source model currently available. The architecture extends previous GLM-5 and GLM-5.1 versions by incorporating multi-head latent attention and sparse attention mechanisms derived from DeepSeek V3.2. The key innovation is IndexShare, a technique for reusing attention indices across model layers. Rather than recalculating sparse-attention indices at every layer, GLM-5.2 computes the full index only once per four layers, allowing subsequent layers to reuse the same token selections. This approach substantially reduces computational overhead for long-context inference while maintaining adaptive attention patterns. Performance benchmarks demonstrate GLM-5.2's capabilities, with a score of 68.8 on the Artificial Analysis Coding Index, significantly outperforming Claude Opus 4.8's 56.7—a margin exceeding ten points for coding tasks.