Zhipu launches GLM-4, claiming near-GPT-4 parity

The Chinese AI company Zhipu launched GLM-4, its fourth-generation foundation model, at the company’s first DevDay conference in Beijing. Chief executive Zhang Peng described it as roughly 60% more capable than its predecessor and said its performance on natural-language benchmarks approached that of GPT-4, with particular strength on Chinese-language tasks.

The model supported a 128,000-token context window — described by the company as maintaining near-perfect accuracy at that length — along with improved multimodal and text-to-image capabilities and faster, cheaper inference than earlier GLM versions. Zhipu also introduced “All Tools,” a feature letting the model plan and execute multi-step tasks autonomously, and “GLMs,” a platform for building customisable agents, both closely mirroring features OpenAI had introduced for ChatGPT months earlier.

Alongside the model, Zhipu announced a fund of roughly $100 million aimed at large-language-model startups, an open-source developer fund, and expanded academic partnerships. The company, spun out of Tsinghua University research and backed by Alibaba and Tencent among other investors, had by this point raised approximately $360 million.

As with other Chinese labs’ GPT-4-parity claims through 2023 and 2024, GLM-4’s benchmark comparisons came from Zhipu’s own reporting rather than an independent third-party evaluation, and the claim of “near-GPT-4” performance could not be independently verified from the announcement alone. The launch nonetheless marked GLM-4 as one of the more capable Chinese frontier models available at the time, and situated Zhipu — alongside Baidu, Alibaba and Moonshot — among the domestic labs racing to match Western frontier models on both raw capability and the surrounding product features that had come to define what a competitive foundation-model release looked like.