GLM-5, z.ai's most recent release, has emerged as the top-performing Open Source LLM on the AIW Models Ranking, marking a new high for open weights models. By the time of the posting of this Story, GLM-5 scores 50 on the Artificial Analysis Intelligence Index, making it the first and only OS model to reach this threshold and outperform other open models such as Kimi K2.5, MiniMax 2.1, and DeepSeek V3.2 across a range of reasoning, coding, and agentic tasks - and only 3 points behind the leader, Claude Opus 4.6!
As usual, the core strength of GLM-5 (and any other OS model) is that it allows developers and researchers to inspect, deploy, and build upon the model without restrictions. The model is exposing a Mixture of Experts architecture, +700 B parameters (with 40 B active parameters on inference) and 200 K token context window, making it great for coding and overall long-horizon tasks. GLM-5 also should be more cost-competitive per token compared to proprietary alternatives like Claude Opus series models, with third-party pricing analyses suggesting significantly lower per-token costs, but with an interesting consideration: since it requires +1 TB of memory to run locally, hardware costs might be higher than just calling Claude's API. Furthermore, GLM-5 remains text-only, lacking multimodal capabilities such as image or video input that other OS models like Kimi K2.5 support, limiting its usefulness in tasks requiring visual understanding.
Still, GLM-5 represents a great achievement in Open Source AI, overtaking previous open leader Kimi K2.5 on many leaderboards and offering performance that closes the gap on proprietary top models. Its open licensing could make it attractive for many applications, but the lack of multimodality and high deployment costs remain practical hurdles that differentiate it from more versatile/accessible offerings in the AI landscape.