Chinese start-up Z.ai, previously known as Zhipu, has released an open-source large language-model family called GLM-4.5 at the World Artificial Intelligence Conference in Shanghai. The flagship system contains 355 billion parameters but activates only 32 billion at inference; a lighter GLM-4.5-Air variant runs on 12 billion active parameters. Both models are distributed under an Apache 2.0 licence and are immediately available for download as well as through Z.ai’s cloud API. Built on a mixture-of-experts architecture, the models provide a 128-kilotoken context window and switch between fast and “thinking” modes to balance latency and accuracy. In third-party tests covering reasoning, coding and autonomous tool use, GLM-4.5 ranks third overall—matching or edging out Claude 4 Sonnet on web tasks and scoring 64.2 % on SWE-bench Verified—while the Air version places sixth. Z.ai says the flagship model can operate on just eight Nvidia H20 chips, hardware designed to comply with U.S. export limits. Z.ai is positioning price as its main differentiator. The company will charge as little as $0.11 per million input tokens and $0.28 per million output tokens for the Air model, and $0.60 / $2.20 for the full-size version. That undercuts DeepSeek’s R1 rates of $0.14 and $2.19 and is well below many U.S. proprietary offerings. Because the weights are open, enterprises can self-host or fine-tune the systems without additional licensing fees. The launch intensifies China’s push to offer low-cost, open-weights alternatives to U.S. leaders such as OpenAI and Anthropic, following recent releases from Alibaba’s Qwen team and Moonshot’s Kimi K2. Z.ai has raised roughly $1.5 billion from Tencent, Alibaba and others, but it also appears on a U.S. entity list that restricts American firms from supplying it. Analysts expect the aggressive pricing and open-source approach to pressure Western model providers to lower costs or accelerate their own open-weights strategies.