Zhipu Releases GLM-5.2: a 744B-Parameter, 1M-Token Coding Model Under a Full MIT License

Chinese AI lab Zhipu (known internationally as Z.ai) has released GLM-5.2, the new flagship in its open-model family — and is putting it out under a fully permissive MIT license. The model is a 744-billion-parameter Mixture-of-Experts that activates about 40 billion parameters per token, trained on 28.5 trillion tokens, and built squarely for long-horizon agentic coding: the multi-step work of writing, running and revising code across an entire project.

Its headline feature is reach. GLM-5.2 ships with a 1-million-token context window (the model id is glm-5.2[1m]) and can return up to 131,072 tokens in a single response — enough to hold a large codebase in view and emit sweeping multi-file changes. The model went live first for subscribers of Zhipu's GLM Coding Plan, with standalone API access, the Z.ai chatbot and the open weights arriving on a staggered timeline over the following week.

The MIT license is the part developers will notice most. Unlike the custom community licenses many "open" models ship with, MIT carries no field-of-use limits, no monthly-active-user threshold and no acceptable-use policy bolted on — making GLM-5.2 genuinely free to embed, fine-tune and resell. That positions it aggressively against the other open-weight heavyweights, alongside Moonshot's Kimi K2.7-Code, MiniMax M3 and DeepSeek V4.

One notable caveat: Zhipu published no benchmark scores at launch. Shipping a flagship model with a marquee 5x context jump and no accompanying evaluation tables is a marketing-first move, and it means any claim that GLM-5.2 "beats" a particular closed model is for now a vendor assertion rather than a measured result. For reference, the prior GLM-5.1 ranked third on Code Arena behind two Claude Opus 4.6 variants and narrowly edged Opus 4.6 on SWE-bench Pro — a strong baseline, but not direct evidence for the new model.

Either way, the release continues a clear 2026 trend: Chinese labs are setting the pace in open weights, shipping permissively licensed, coding-focused models with frontier-scale context windows faster than their Western counterparts will open-source anything comparable.