GLM 5.3 Flash

GLM 5.3 Flash is the first natively multimodal model in Z.ai's GLM-5 series, a mixture-of-experts model with 320 billion total parameters and just 18 billion active per token, accepting text and image input and returning text. Built on a newly trained base with a 30T-token multimodal corpus, it introduces a hybrid architecture that interleaves three linear attention layers with one DeepSeek sparse attention layer, plus Manifold-Constrained Hyper-Connections (mHC), to carry a 1 million token context at a fraction of the compute. Z.ai reports it outperforms GLM 5.2 across benchmarks and real-world workloads at one-tenth the price while approaching Claude Opus 4.8 on coding and agentic tasks, making it well-suited to high-throughput agents, document and screenshot understanding, and cost-sensitive production deployments.