技术规格

GLM-5

GLM-5 is a flagship multimodal foundation model developed by Z.ai, designed for complex systems engineering and long-horizon agentic workflows. Utilizing a Mixture-of-Experts (MoE) architecture, the model scales to 744 billion total parameters with approximately 40 billion parameters activated per token. This design facilitates high-capacity reasoning and specialized knowledge retrieval while maintaining the computational efficiency required for large-scale deployment. The model is trained on a massive 28.5 trillion token corpus, emphasizing high-quality code, technical documentation, and reasoning-dense data to support professional-grade software development and autonomous problem-solving.

Technically, GLM-5 introduces several architectural innovations, most notably the integration of DeepSeek Sparse Attention (DSA). This mechanism optimizes the standard attention block by dynamically allocating computational resources, which significantly reduces the memory and compute overhead associated with processing long sequences. Additionally, the model leverages an asynchronous reinforcement learning infrastructure known as 'slime' during post-training. This framework decouples generation from training to improve iteration throughput, allowing the model to learn effectively from complex, multi-step interactions and dynamic environments.

Optimized for long-context stability, GLM-5 supports a context window of up to 204,800 tokens and is capable of generating up to 128,000 tokens in a single output. Its operational capabilities include advanced tool-use, real-time streaming, and structured output across frontend, backend, and data processing tasks. The model is released with open weights under the MIT License, enabling researchers and developers to perform local serving, fine-tuning, and integration into diverse agentic frameworks without vendor lock-in.