Z.ai Releases GLM-5.3-Flash Under MIT: 320B Total, 18B Active, First Natively Multimodal GLM-5 Model

Z.ai published GLM-5.3-Flash on Hugging Face on August 26, 2026. The model runs 320B total parameters with 18B active, and the company says it beats GLM-5.2 at one-tenth the price while approaching Claude Opus 4.8 on coding and agentic benchmarks. The license is MIT.

GLM Z-ai open-weight multimodal LLM China

Z.ai (zai-org) published GLM-5.3-Flash on Hugging Face on August 26, 2026¹. It is the first natively multimodal model in the GLM-5 series, configured with 320B total parameters and 18B active¹. The company states that it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, and that it approaches Claude Opus 4.8 on coding and agentic benchmarks¹. Both are Z.ai’s own claims rather than independently reproduced results.

The architecture was rebuilt starting from a newly trained base model, introducing for the first time in the GLM series a hybrid of sparse and linear attention, which the company says cuts long-context serving costs while keeping long-context accuracy intact¹. It also adopts Manifold-Constrained Hyper-Connections (mHC) for scaling efficiency, which together with a 30-trillion-token multimodal pre-training corpus is meant to deliver more capability per unit of compute¹.

The license is MIT, with local deployment supported through SGLang, vLLM, TokenSpeed, and KTransformers alongside Z.ai’s own API¹. The company also released GLM-5.3 on August 14, when it said it would hold the weights back until safety evaluation was complete.

Sources

GLM-5.3-Flash - Z.ai (zai-org) official model card (August 26, 2026)