GLM-4.5

GLM-4.5 is the release where Z.ai stopped being a domestic curiosity and started being quoted in the same breath as Western frontier models. Published in late July 2025, it was built from the start as a foundation for agents rather than for chat: tool invocation, web browsing, software engineering and front-end work are the four jobs the company names in its own documentation. The architecture is mixture-of-experts. Z.ai reports 355 billion total parameters with 32 billion active per forward pass; the safetensors index on Hugging Face adds up to 358.3 billion, the usual small gap between a rounded headline figure and what the weight files actually contain. Pretraining ran on 15 trillion tokens of general data, followed by fine-tuning on code, reasoning and agent-specific sets, then reinforcement learning. The context window is 128,000 tokens with up to 96,000 tokens of output — a large answer budget for its generation. It was also the first GLM with hybrid reasoning: Thinking Mode for hard problems and tool chains, Non-Thinking Mode for instant replies, switched by a single API parameter. The line shipped as a family — the standard model, the lighter Air, the fast X and AirX tiers, and a free Flash variant — a pattern Z.ai has repeated with every generation since. What makes it worth reading today is the price list. GLM-4.5 still stands at $0.60 per million input tokens and $2.20 per million output, exactly what Z.ai charges for GLM-4.6 and GLM-4.7. Three generations, one price: the company competes on capability, not on discounting its back catalogue. The weights are on Hugging Face under an MIT licence, one of the most permissive terms any model of this size has been released under.

▸Videos

No videos yet.

▸Reviews