Zai releases GLM-5.3-Flash, a 320B multimodal model with sparse-linear attention

Zai has introduced GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. It features a hybrid architecture combining sparse and linear attention to reduce long-context serving costs while maintaining precise capabilities.

• The model has 320B total parameters with only 18B active parameters.
• It outperforms GLM-5.2 across benchmarks at one-tenth the price.
• Performance approaches Claude Opus 4.8 on coding and agentic benchmarks.
• Uses Manifold-Constrained Hyper-Connections (mHC) for improved scaling efficiency.
• Trained on a 30T-token multimodal pre-training corpus.

The model is available via Z.ai API Platform and supports deployment with SGLang, vLLM, TokenSpeed, KTransformers, and HLE.