glm-4.7

GLM-4.7 is a large-scale Mixture-of-Experts model from Zhipu AI designed from the ground up for intelligent agent applications. With 355 billion total parameters and 32 billion active parameters per forward pass, it achieves computational efficiency by activating only relevant expert pathways for each task, much like biological neural systems engage specific regions based on demand. This architecture enables it to handle extremely long contexts up to 200,000 tokens and generate up to 128,000 tokens in a single output pass, allowing developers to produce entire software frameworks in one go. The model explicitly supports "thinking before acting" reasoning modes, enabling it to deliberate on complex problems before responding, and it excels particularly at multilingual agentic coding, terminal-based operations, and tool-using workflows.

The GLM-4.7 release marks a significant leap over its GLM-4.6 predecessor across nearly every benchmark category, with gains including 5.8% improvement on SWE-bench coding tasks, 12.9% on multilingual coding evaluations, 16.5% on Terminal Bench 2.0, and 12.4% on the HLE Humanity's Last Exam benchmark. Tool-using capabilities show particularly notable advancement through improved performance on τ²-Bench and web browsing benchmarks. The model also introduces enhanced "Vibe Coding" abilities for frontend development, producing cleaner, more modern webpage layouts and better-sized presentation slides. Being open-weight on Hugging Face means developers can run it locally or deploy it through various managed services, giving the open-source community direct access to a model that competes with proprietary alternatives in coding, reasoning, and agentic workflows.