GLM5 Emerges as Groundbreaking Open-Weight LLM, Challenges Frontier Models in Code and Agentic Tasks

Z.ai has unveiled GLM5, an open-weight language model that is significantly impacting the competitive landscape, particularly for software development tasks. Positioned as the most capable open-weight model to date, GLM5 demonstrates performance akin to established closed-weight models such as Anthropic’s Opus 4.5 and Google’s CodeX 5.2, while notably outperforming Gemini 3 Pro. The model is available under an MIT license and is accessible via a dedicated desktop application, though this requires Z.ai hosting. Its architecture scales to 744 billion parameters (with 40 billion active using a Mixture of Experts approach) and leverages Deepseek sparse attention alongside an innovative asynchronous reinforcement learning infrastructure, ‘Slime’, for enhanced fine-tuning efficiency.

GLM5’s benchmark results are particularly compelling, showcasing leading scores in areas like Humanity’s Last Exam (with tools) and a record-breaking performance in BrowseComp, a browser usage benchmark. Crucially for developers, it achieved near parity with top models in SWE-bench Verified and Terminal Bench. The model exhibits impressive long-running task capabilities, successfully completing an hour-long codebase migration without issues, a feat previously challenging for any LLM, let alone an open-weight one. Furthermore, GLM5 sets a new low for hallucination rates (around 30%) by frequently abstaining rather than providing incorrect information, a significant improvement over models like Gemini 3 Pro (88%). While it currently lacks multimodal image input, its cost-effectiveness — estimated at 3-6x cheaper than leading closed-weight models for similar tasks — combined with its robust performance, positions GLM5 as a compelling alternative for engineers and organizations seeking advanced AI capabilities at a more accessible price point.