zai-org/GLM-5.3-Flash (huggingface.co)

The release of GLM-5.3-Flash marks a significant advancement in the GLM series as it becomes the first natively multimodal model, boasting a total of 320 billion parameters, with 18 billion active parameters, allowing it to outperform its predecessor, GLM-5.2, on various benchmarks at a fraction of the cost. This model's architecture introduces a hybrid approach that combines sparse and linear attention, effectively reducing the computational costs associated with long-context processing while retaining its accuracy. Its new training strategy leverages a massive 30 trillion-token multimodal pre-training corpus to enhance performance, particularly in coding and agent tasks, where it approaches the capabilities of Claude Opus 4.8. For the AI/ML community, these developments are crucial not only for operational efficiency—achieving more intelligence with less computational power—but also for expanding the applicability of the GLM series across various frameworks. The model supports deployment in several environments including HLE with tools, NL2Repo, and DeepSWE, among others. With enhanced evaluation metrics and context management strategies in place, GLM-5.3-Flash positions itself as a robust solution for researchers, emphasizing the ongoing trend toward more efficient and capable AI systems.