GLM-5: Z.ai Releases a 744B-Parameter Open-Weight Model

A 744-billion-parameter mixture for agent work

Z.ai released GLM-5 as a 744-billion-parameter mixture-of-experts model with 40 billion active parameters, up from GLM-4.5's 355 billion total and 32 billion active parameters. The technical report says the model has 256 experts and 80 layers, and that its base-model training used 28.5 trillion tokens. [2][1]

GLM-5 added DeepSeek Sparse Attention during continued pre-training. The authors describe it as selecting important tokens dynamically instead of applying dense attention to every token pair, an efficiency measure intended to preserve long-context performance while reducing attention computation. They also describe an asynchronous reinforcement-learning system that separates generation from training and agent-RL work on longer interactions. [2][1]

Atlas interpretation: The important distinction is between the model's total footprint and the computation it activates for a request. A 744-billion-parameter headline did not mean every parameter was active at once, but it still put a 40-billion-active-parameter open-weight model in a deployment class far beyond a typical single-machine download. Sparse attention and the mixture are attempts to make that scale usable for the long-context, tool-using jobs Z.ai was targeting. [2][1]

Weights available alongside the API

The release made GLM-5's model repository available on Hugging Face under an MIT license. The pinned model card gives examples for loading it through Transformers and serving it locally with vLLM, SGLang or Docker, while also pointing readers to Z.ai's API platform. That is an open-weight release with multiple deployment paths, rather than a model reachable only through Z.ai's hosted service. [1]

Atlas interpretation: Open weights make the model portable, not lightweight. The documentation's production examples use eight-way tensor parallelism and specialized Hopper or Blackwell containers, which is a reminder that permission to download a frontier-sized model and the hardware needed to serve it are different forms of access. [1]

A launch table, not an independent verdict

Z.ai's model card reported GLM-5 at 77.8 on SWE-bench Verified, 56.2 or 60.7 on two Terminal-Bench 2.0 variants, 43.2 on CyberGym and 30.5 on Humanity's Last Exam. Its table compared those results with GLM-4.7, DeepSeek-V3.2, Kimi K2.5, Claude Opus 4.5, Gemini 3 Pro and GPT-5.2. The accompanying report similarly characterized the release as an open-weights leader and comparable to selected closed models on its reported suite. [1][2]

Atlas interpretation: Those are developer measurements, not independently replicated results. The card does disclose material evaluation choices, including tailored prompts, different context limits, tool settings and a GPT-5.2 judge for some tasks; it also notes that some comparator scores cover a full set while GLM-5's reported score may use a text-only subset. The table is useful evidence of what Z.ai tested and claimed at launch, but it cannot by itself settle a general ranking across models or real-world agent reliability. [1][2]