TL;DR: Zai has open-sourced GLM 5 (previously Pony Alpha), a 744B parameter Mixture-of-Experts model designed for long-horizon agentic and systems engineering tasks.
Summary: GLM 5 is a Mixture-of-Experts model with 744B total parameters and 40B active parameters, trained on 28.5T tokens using DeepSeek Sparse Attention (DSA) and an asynchronous reinforcement learning infrastructure. Released under an MIT license, the model supports native FP8 precision and is integrated with vLLM and SGLang for local deployment.
Why it matters: It gives indie builders access to a highly competitive, MIT-licensed open model optimized for complex agentic engineering and coding tasks. Developers can run it locally using vLLM/SGLang or access weights directly via Hugging Face and OpenRouter.
Source: t.me