Introducing GLM-5.3-Flash: a new multimodal model from zhipu AI with significant cost reduction

Zhipu AI has officially launched and open-sourced GLM-5.3-Flash, a native multimodal model in the GLM-5 series, with a total parameter count of 320 billion and an active parameter count of 18 billion. This model was previously tested under the anonymous name Ox-Alpha, which gained significant popularity on OpenCode and OpenRouter, achieving high traffic volumes supported entirely by domestic chips. GLM-5.3-Flash outperforms GLM-5.2, scoring 57 on the Artificial Analysis Intelligence Index (AA), placing it within the global advanced model capability range and matching the score of Anthropic's Claude Opus 4.8. In Z.ai's internal benchmark tests, its coding performance is comparable to that of Claude Opus 4.8. The model is priced at one-tenth of GLM-5.3, with a limited-time discount bringing it down to one-twentieth of GLM-5.3, which is one-fortieth of Opus 4.8's price. Zhipu AI claims that GLM-5.3-Flash surpasses GLM-5.2 in performance despite having half the parameters and a tenth of the cost. This achievement is attributed to a new model architecture designed for low cost. GLM-5.3-Flash has a parameter count similar to GLM-4.5 (355B vs. 320B), but with nearly halved active parameters (32B → 18B) and layers (92 → 45). Combined with the latest 30T Token multimodal pre-training data, it achieves stronger performance with fewer computational resources. The model also features several architectural upgrades, including the first hybrid sparse attention and linear attention architecture in an open-source model. This maintains precise long-context capabilities while significantly reducing service costs. Additionally, it uses manifold-constrained hyper-connections (mHC) to enhance scaling capabilities. Zhipu AI highlights that GLM-5.3-Flash integrates visual capabilities natively into coding loops, allowing the model to observe interfaces, render results, and receive feedback to continuously test and improve. It supports tasks such as frontend development, game building, Blender 3D scenes, and real-world environment operations driven by BUA and CUA. The model has also expanded into office, financial research, and professional document scenarios. It can autonomously break down complex tasks, call appropriate tools, check, and optimize outputs, completing the full workflow from research analysis to model building and final delivery in formats like PPTX, PDF, DOCX, and XLSX. GLM-5.3-Flash is now open to the global community, integrated with platforms like ZCode, and included in the GLM Coding Plan, which offers limited daily experience cards. API access is also available. Zhipu AI has reset the overall quota for the GLM Coding Plan, and all users can now see their usage credits fully replenished in their accounts. For more information, visit the BigModel Open Platform, Z.ai, GLM Coding Plan, and HuggingFace links provided.