Zhipu AI open-sources GLM-5.3-Flash with 320 billion parameters at one-tenth the cost of GLM-5.2

Dongcha Beating AI News: After claiming Ox Alpha during the day, Zhipu officially open-sourced GLM-5.3-Flash in the evening; model weights are now available on Hugging Face under the MIT license. With a total of 320B parameters, it activates only 18B at a time, natively supports text, images, and video, and is the first natively multimodal model in the GLM-5 series. According to the official team, GLM-5.3-Flash outperforms GLM-5.2 overall, with programming and agent evaluation scores approaching Claude Opus 4.8, yet at just one-tenth the cost of GLM-5.2. It also introduces a new base architecture, incorporating for the first time in the main GLM series a hybrid sparse and linear attention mechanism, pre-trained on 30T multimodal tokens. Previously, community tests on Ox Alpha’s DeepSWE small-sample benchmark briefly reached 80%, but that was based on only 10 questions; when expanded, performance dropped to around 63%, and testers have since corrected their initial claims. This result still places it in the top tier, though not as dramatically high as the initial 80%. With weights now open-sourced, developers can directly deploy the model using frameworks like vLLM, SGLang, and KTransformers without relying on anonymous model APIs.