GLM-5 Becomes Top-Ranked Open-Weight Code Model
Zhipu AI and Tsinghua University have released GLM-5, a new open-weight model that now ranks first for both code and text generation on major benchmarks like Artificial Analysis and LMArena. The model reportedly outperforms some closed-source models, including Claude and GPT variants, on certain agentic tasks. This release reinforces the trend toward capable, open-source alternatives for AI development.
- The model utilizes a Mixture-of-Experts (MoE) architecture with 744 billion total parameters, but only 40 billion are active during any given inference, a design intended to balance high performance with manageable computational costs. It was trained on 28.5 trillion tokens of data. - On the Vending Bench 2 benchmark, which simulates running a business for a year to test long-term planning, GLM-5 achieved a final balance of $4,432, ranking first among open-source models. It also scored 77.8% on the SWE-bench Verified coding evaluation. - A key aspect of its development is the use of domestic Chinese hardware; Zhipu AI has stated that GLM-5 can be deployed using chips from companies like Huawei, Moore Threads, and Cambricon, reducing reliance on Nvidia. - The model incorporates a novel asynchronous reinforcement learning (RL) infrastructure called "slime" to improve post-training efficiency and DeepSeek Sparse Attention (DSA) to reduce deployment costs while maintaining a long-context capacity of over 200,000 tokens. - GLM-5 is notable for its large output capacity, able to generate up to 128,000 tokens in a single pass, a significant increase over the typical 4,000 to 16,000 token limits of many other models. - Its model weights are released under the permissive MIT license, allowing for broad commercial use. Access via third-party APIs can be 5 to 10 times cheaper than proprietary competitors like Claude Opus 4.5. - On the Artificial Analysis Intelligence Index, GLM-5 was the first open-weight model to achieve a score of 50, significantly closing the performance gap with leading proprietary models.
