Z.ai's GLM-5 Sets New Open-Source AI Benchmark
Z.ai's new GLM-5 model has achieved the highest-ever "Intelligence Index" score for an open-weights model. At more than double the size of its predecessor, the model's performance in reasoning and multimodal tasks demonstrates that open-source LLMs are increasingly competitive with proprietary alternatives.
- GLM-5 is a Mixture-of-Experts (MoE) model with 744 billion total parameters, of which 40 billion are active for any given token. This is more than double the 355 billion total parameters of its predecessor, GLM-4.5. - The model was trained on 28.5 trillion tokens of data and utilizes DeepSeek Sparse Attention (DSA) to efficiently manage a context window of over 200,000 tokens. - On the SWE-bench Verified benchmark for software engineering tasks, GLM-5 achieves a score of 77.8%, surpassing the performance of proprietary models like Gemini 3 Pro. - A key architectural innovation is a new asynchronous reinforcement learning infrastructure called "slime," which improves the efficiency of post-training workflows. - According to reports, the model was trained entirely on a cluster of Huawei Ascend AI chips, demonstrating a move away from dependency on NVIDIA hardware for training frontier models. - For self-hosting, the model's weights require approximately 1.5 TB of memory in their native BF16 precision. Z.ai provides support for deployment on various non-NVIDIA chips, including those from Huawei, Moore Threads, and Cambricon. - The model's weights are available on platforms like Hugging Face and ModelScope under the permissive MIT License, allowing for commercial use.
