GLM-5

GLM-5 is a 744B parameter open-source AI model specifically engineered for complex systems engineering and long-horizon agentic tasks. As a complex systems engineering AI, it targets developers and researchers who demand robust reasoning, coding, and autonomous planning capabilities. By scaling from its predecessor's 355B to 744B parameters and pre-training on 28.5T tokens, GLM-5 delivers significant improvements across academic and real-world benchmarks. Its core value lies in achieving best-in-class performance among open-source models on reasoning, coding, and agentic evaluations, while narrowing the gap with proprietary frontier models like Claude Opus 4.5 and GPT-5.2.

Traditional open-source models often falter when faced with long-horizon tasks that require sustained planning and resource management over extended periods. They also struggle with complex systems engineering, where multiple interdependent components must be coordinated. This pain point is especially acute for software engineers and AI researchers who need to deploy models for tasks like full-stack development, multi-agent collaboration, or business simulations. GLM-5 addresses these limitations through its large-scale architecture and advanced post-training techniques, enabling it to handle tasks that run for hundreds of steps or simulate entire fiscal years with high competence. For example, on Vending Bench 2, GLM-5 manages a simulated vending machine business over a one-year horizon, ending with a $4,432 account balance that approaches frontier model performance.

One of GLM-5's most impactful features is its integration of DeepSeek Sparse Attention, or DSA. This technique reduces the computational cost of deploying the model by sparsifying attention mechanisms, meaning only the most relevant tokens are processed at each step. Crucially, DSA preserves the model's long-context capacity, allowing it to handle sequences of up to 200K tokens without degrading performance. For users, this translates to lower hardware requirements and reduced costs when running inference, making a 744B model practical for everyday use. The benefit is particularly evident in agentic workflows where the model must maintain coherence across numerous interactions, such as browsing the web or executing multi-step code repairs.

GLM-5 also introduces the slime asynchronous reinforcement learning infrastructure, a custom framework developed to overcome the inefficiencies of scaling RL for large language models. Slime parallelizes training across multiple nodes with asynchronous parameter updates, dramatically improving training throughput and enabling more granular post-training iterations. This means the model can be fine-tuned on specific tasks—like coding or reasoning—with greater speed and precision. The result is a model that not only understands complex instructions but also refines its responses through iterative learning, as seen in its top-tier performance on SWE-bench Verified (77.8%) and Terminal-Bench 2.0 (56.2%), where it outperforms many open-source competitors.