GLM-5.3-Flash Officially Open-Sourced, 320B Parameters Activate 18B, Priced at One-Tenth of GLM-5.2

Zhipu has officially open-sourced GLM-5.3-Flash, with model weights now available on Hugging Face under the MIT license. The model has a total of 320 billion parameters, but only activates 18 billion at a time, supporting text, images, and videos, making it the first native multimodal model in the GLM-5 series. The official statement indicates that GLM-5.3-Flash outperforms GLM-5.2 overall, with programming and agent evaluations approaching those of Claude Opus 4.8, and is priced at only one-tenth of GLM-5.2. The new model introduces a sparse attention and linear attention mixed architecture, and is pre-trained using 30 trillion tokens of multimodal data. Community tests on Ox Alpha's DeepSWE small sample reached as high as 80%, but scores fell to about 63% after expanding the sample size. After the weights are open-sourced, developers can deploy using frameworks like vLLM, SGLang, and KTransformers without relying on the API of anonymous models.