Z.ai Releases GLM 5.2 Open Model with 1M Context and MIT License

Z.ai has released GLM 5.2, a 753 billion parameter open model designed for long-horizon coding and engineering tasks, according to a company blog post. The model supports a 1 million token context window and is available immediately through the Z.ai API, Hugging Face, and other coding platforms.

GLM 5.2 introduces an architectural optimization called IndexShare, which reuses the same indexer across every four sparse attention layers to reduce computation by nearly three times at maximum context length. The model also includes an improved Multi Token Prediction layer for speculative decoding, increasing acceptance length by up to 20 percent. Users can select between two reasoning effort levels, High and Max, to balance performance and latency.

Benchmarks show that GLM 5.2 performs above most open models and close to proprietary ones such as GPT 5.5 from OpenAI and Claude Opus 4.8. It scored 62.1 on SWE-bench Pro, 74.4 on FrontierSWE, and 81.0 on Terminal Bench 2.1, ranking as the strongest open model across several coding tasks.

The model’s weights are available under an MIT open-source license, allowing unrestricted use, modification, and commercial deployment. GLM 5.2 can be used locally through frameworks such as transformers and vLLM. Z.ai also offers the model through its Coding Plan, which integrates with third-party coding tools and includes usage tiers starting at $12.60 per month.