Zhipu, one of China's largest artificial intelligence labs, announced GLM 5.2 today with a 1 million token context window — roughly the text of a stack of novels the model can read at once — and a claim of leadership on long coding tasks.
Zhipu AI, one of China's largest artificial intelligence labs, says it has shipped GLM-5.2, a new open-weight language model it calls its "most capable" yet. The model is being framed as a counter to recent restrictions on other frontier AI systems, and the company is using the launch to stake out a position on open-source AI. None of the headline capability numbers has been independently verified, the model card is not yet public, and the weights are not yet downloadable.
The announcement came in an X post from Zhipu CEO Jie Tang on June 13, 2026, which called GLM-5.2 "Fully Open" and declared that "Frontier Intelligence Belongs to Everyone." Tang said GLM-5.2 has a "truly usable 1 million token context window," five times the 200,000-token limit of its predecessor GLM-5.1, and that it "maintains a continuous lead in the independent completion of long-horizon tasks." He also called the model "the main engine for creating the strongest domestic coding model," an implicit comparison to other Chinese AI systems like Alibaba's Qwen and DeepSeek. A token is roughly four characters of English text, so a 1 million token context window is large enough to hold a stack of novels in a single prompt.
The claims sit inside a political frame. Tang described the release as a response to the "sudden restriction of certain frontier models" and "external blockades and restrictions," language that echoes the U.S. export-control debate over advanced AI chips and the recent curtailment of Chinese users' access to some Western AI services. The framing of GLM-5.2 as "radical openness" is a deliberate counterpoint, but it is the company's own rhetoric, not an independent assessment.
The verifiable record, for now, is the prior lineage. Zhipu published the GLM-5 technical report on arXiv in February 2026, describing a 744-billion-parameter mixture-of-experts model with about 40 billion parameters active per pass, trained on roughly 28.5 trillion tokens and incorporating DeepSeek Sparse Attention. The model is hosted in the official zai-org/GLM-5 GitHub repository and licensed under MIT. The GLM-5.1 model card on Hugging Face lists a 200,000-token context window with a 128,000-token maximum output and benchmarks that place it ahead of comparable models on SWE-Bench Pro, NL2Repo, and Terminal-Bench 2.0, with Z.ai's own GLM-5.1 API documentation describing it as designed for "up to ~8 hours of autonomous work."
That lineage matters because it is the basis for the GLM-5.2 claim, and because the MIT license on GLM-5 and 5.1 has been a major reason Western developers have been willing to build on the model. Whether GLM-5.2 ships under the same terms, or under something more restrictive, is not yet known. The Hugging Face page for the model returns a 404 as of this writing, and the Zhipu organization's public model list is empty.
The rollout itself is staged. Tang said GLM-5.2 would become available to all GLM Coding Plan subscribers, including Lite, Pro, and Max tiers, on the evening of June 13, with a public API expected to follow "next week." A downloadable weights release, which would let independent researchers run and benchmark the model, has not been announced. Until the model card, the license, and the weights are public, every benchmark number circulating about GLM-5.2 is a Zhipu number, not an independent one.
That leaves developers and policy watchers in a familiar waiting pattern: watch for the model card to post, watch for the first independent benchmark run on the public API next week, and read the political framing as political framing. The argument over whether open weights are good for AI safety, or whether the U.S. restrictions Tang alluded to are good or bad policy, will continue either way. The new question is whether GLM-5.2, when it is actually downloadable, will perform as the CEO's post claims it does.