**["Bull Arrival" Model GLM-5.3-Flash Officially Open-Sourced: 320B Parameters, Only 18B Activated, Price is One-Tenth of GLM-5.2]** Beating AI News Flash: After Ox Alpha was claimed during the day, Zhipu officially open-sourced GLM-5.3-Flash in the evening. The model weights are now available on Hugging Face under an MIT license. It has a total of 320B parameters, but only 18B are activated at a time. It natively supports text, images, and videos, making it the first native multimodal model in the GLM-5 series. According to the official announcement, GLM-5.3-Flash outperforms GLM-5.2 in overall performance, with programming and agent evaluations approaching Claude Opus 4.8, yet its price is only one-tenth of GLM-5.2. The model also introduces a new foundational architecture, incorporating sparse attention and linear attention hybrid structures into the main GLM series for the first time. It was pre-trained using 30T tokens of multimodal data. Previously, the community's small-sample DeepSWE test on Ox Alpha achieved an impressive 80% score, but that test only consisted of 10 questions. After expanding the sample size, the score dropped to approximately 63%, and testers voluntarily corrected their initial claims. While this result still places it in the top tier, it is not as exaggerated as the initial 80%. With the weights now open-sourced, developers can directly deploy the model using frameworks like vLLM, SGLang, and KTransformers, without needing to rely on anonymous model APIs. [Original Link]