Zhipu AI

Zhipu AI (智谱AI) is a Chinese AI company founded in 2019 as a spinoff from Tsinghua University, one of China's top research institutions. It develops the GLM (General Language Model) family and is one of China's most prominent and well-capitalized AI labs, competing at the frontier alongside DeepSeek, Alibaba's Qwen, Moonshot AI, and MiniMax. In January 2026 it became the first large language model company to go public, listing on the Hong Kong Stock Exchange and raising roughly 558 million dollars. It has leaned on those public markets hard since. A follow-on share sale in July 2026 raised about 4 billion dollars, and on September 13, 2026 the company placed roughly 2 billion dollars of new shares alongside a convertible bond sale of about 3 billion dollars, taking the total raised since listing to roughly 9.6 billion dollars. The new shares priced at a 10 percent discount to the previous close, and the company earmarked about 60 percent of the net proceeds for research on next-generation models and self-training systems, with a further 15 percent for expansion.

Its flagship line is GLM-5, released in February 2026, followed by the coding-first GLM 5.2 in June 2026 and GLM-5.3 in August 2026 — a mixture-of-experts line with 744 billion total parameters and a context window that reached 1 million tokens. Licensing now varies by model rather than by generation, and the GLM-5.3 generation is where that became explicit: GLM-5 and GLM 5.2 shipped MIT-licensed open weights, GLM-5.3-Flash followed under a plain MIT license on August 27, 2026, and the flagship GLM-5.3 weights arrived a day later under a custom GLM-5.3 License — an MIT-shaped grant that adds a Z.ai security review for any model-as-a-service operator whose revenue passes 10 billion dollars over twelve consecutive months. What sets GLM-5 apart is how it was built: entirely on Huawei Ascend chips using the MindSpore framework, with zero NVIDIA dependency — a proof point that frontier-scale models can be trained on fully domestic Chinese hardware despite US export controls. Zhipu now operates its consumer and international platform under the Z.ai brand.

Zhipu AI benefits from its deep connection to Tsinghua University's research ecosystem, which produces many of China's top AI researchers and engineers. The company differentiates through that academic pedigree, its NVIDIA-free training stack, and a consistently permissive open-weights strategy — GLM-5 and GLM 5.2 are released under the MIT license and have been downloaded millions of times. That posture met its first real test with GLM-5.3 in August 2026: Z.ai shipped the model through its API but held the open weights back for roughly two weeks of safety evaluation, saying cyber capability had grown faster than expected as post-training scaled, and it did not name a license at launch. It is a rare instance of an open-weights lab delaying a release on capability-risk grounds rather than shipping first. The hold ended on August 28 with the weights published rather than withheld, but not on the line’s usual terms — the flagship is the first GLM model to carry conditions beyond MIT.

On September 10, 2026 Anthropic named Zhipu in a threat intelligence report as running an unauthorized distillation campaign against Claude, attributing more than 3.4 million exchanges to it over seventeen days in June and July 2026. The detail worth carrying is what Anthropic says happened next: after its cyber safeguards degraded the attacks, Zhipu gave up on targeting the Fable line and switched to models it assessed had weaker safeguards — Opus 4.6 and the leading model of another US lab. If accurate, that is a useful illustration of how safety investment actually behaves in a competitive market: hardening one model does not stop an adversary, it redirects them to a softer target. Zhipu had not publicly responded at publication, and Anthropic competes with it directly.