Z.ai opens GLM-5.2
On 16 June 2026 Z.ai released GLM-5.2, a flagship model with a context of one million tokens, and published its weights on Hugging Face under the MIT licence with, in its words, no regional limits.
Why it matters
A model with a million-token context, suited to long agentic tasks, became available to anyone to run and fine-tune under the most permissive of the common licences. An editorial assessment.
By the Z.ai blog, the model describes itself as a flagship for long-horizon tasks, has several thinking-effort levels and a mechanism called IndexShare that reuses one indexer for every four sparse-attention layers and, by the company's statement, cuts per-token computation 2.9 times at a million-token context. The repository zai-org/GLM-5.2 was created on 16 June 2026; by its safetensors metadata it holds about 753 billion parameters. What the record does not claim. The benchmark table in the blog is the company's own; we did not reproduce the comparison with other labs' models. The figures '744 billion parameters, 40 active' that circulate in secondary summaries are not on the pages read. The API documentation page the report named has no date and no licence.