GLM-5 Open Source: Zhipu Shifts from Code Generation to Long-Horizon Systems Engineering
Zhipu released GLM-5 in February 2026, open-sourcing weights under the MIT license and providing APIs both domestically and internationally.
GLM-5 repositions the model from Vibe Coding to Agentic Engineering, emphasizing complex systems engineering, long-horizon tasks, and sustainable optimization, marking a major upgrade in Zhipu's global developer roadmap.
GLM-5 has 744B total parameters and 40B activated parameters, with pre-training data expanded to 28.5T tokens. It introduces DeepSeek Sparse Attention and uses an asynchronous RL infrastructure called slime to improve post-training efficiency.
Chinese open models continue to approach global frontiers in scale, licensing, and agent engineering goals, while driving adaptation of domestic chips, inference frameworks, and coding tools.
Enterprises should prioritize engineering tasks that fully cover testing, fixing, and delivery, evaluating value based on hourly success outcomes; code generation volume should only be a secondary metric.
Observe real long-horizon task completion rates, MIT ecosystem adoption, non-NVIDIA deployment efficiency, and API high-concurrency stability.