Z.ai Launches GLM-5.3-Flash, a 320B MoE Model With 1M Context
Z.ai has released GLM-5.3-Flash, a new large language model built on a mixture-of-experts (MoE) architecture, according to Marktechpost. The model carries 320 billion total parameters, of which roughly 18 billion are active per token, a design intended to balance model capacity with inference efficiency.
Per Marktechpost, GLM-5.3-Flash is described as natively multimodal, meaning it is built to process multiple input types as part of its core architecture rather than through bolted-on extensions. The report also highlights a context window of up to 1 million tokens, allowing the model to handle very long documents, codebases, or conversations in a single pass without needing to chunk inputs.
The MoE approach used in GLM-5.3-Flash routes each input through a subset of specialized sub-networks, or experts, rather than activating the entire parameter set. This lets the model maintain a large overall parameter count for capacity while keeping the computational cost of each inference call closer to that of a much smaller dense model, since only the 18 billion active parameters are engaged per token.
Marktechpost’s report frames the release as part of Z.ai’s ongoing GLM model series, positioning the company among labs pursuing large-scale, efficient, and long-context multimodal systems. The combination of a large total parameter count, a smaller active parameter footprint, and an extended context window reflects a broader industry trend toward models that aim to deliver strong performance while managing compute and memory demands during deployment.
Further technical details, including specific benchmark results, training data composition, and availability of model weights, were not fully specified in the available report. AI NewPulse will continue to track official documentation from Z.ai as it becomes available.
Based on reporting by www.marktechpost.com.
