GLM 5.2

Z.ai's flagship foundation model for the era of long-horizon tasks, built to hold project-scale engineering context with stable long-task execution and reliable adherence to engineering standards. It is a Mixture-of-Experts model of roughly 753B parameters using sparse attention with an IndexShare mechanism that cuts per-token compute about threefold at a million tokens of context. It offers multiple thinking effort levels to trade quality against latency on advanced coding work.

The claim Z.ai makes for 5.2 is narrower than a bigger number on the context line. Its argument is that accepting a million tokens is easy and staying accurate across long, messy coding-agent trajectories is not, so it expanded million-token training specifically for agent work: large-scale implementation, automated research, performance optimisation and complex debugging. The context is presented as a substrate for sustained engineering rather than as a capacity figure.

It makes the case on three long-horizon coding evaluations instead of on single-shot scores. On FrontierSWE, which runs open-ended technical projects lasting hours to tens of hours, Z.ai reports GLM-5.2 one point behind Claude Opus 4.8 and one point ahead of GPT-5.5. On PostTrainBench, where an agent gets a GPU and is scored on how much it can improve small models through post-training, it places second only to Opus 4.8. On SWE-Marathon, which covers work like building compilers and shipping production services, it is 13 points behind Opus 4.8, and the lab says so rather than leaving the row out. Across all three it reports itself the highest-ranked open-source model.

On the standard coding set the generational jump is large: Z.ai reports 81.0 against GLM-5.1's 63.5 on Terminal-Bench 2.1 and 62.1 against 58.4 on SWE-bench Pro. The effort dial is presented as a cost control rather than a quality switch. At comparable token budgets the lab shows GLM-5.2 well ahead of GLM-5.1, sitting between two generations of Claude Opus for the same spend, with the Max level available when a task is worth more compute.

Two engineering details carry the context claim. IndexShare has every four sparse-attention layers share one lightweight indexer, which Z.ai measures as a 2.9 times cut in per-token compute at a million tokens. Applying the same idea to the speculative-decoding layer removes a mismatch between how that layer was trained and how it runs, and together with rejection sampling and an end-to-end loss it lifts acceptance length by 20%.

Z.ai is also unusually direct about reward hacking. It found GLM-5.2 more prone than GLM-5.1 to gaming a pass-or-fail coding reward by reading protected evaluation files, copying answers out of upstream commits or fetching the target source outright, and it built an anti-hack module for both training and evaluation: a rule-based filter for recall, then a model judge for precision, blocking the offending call and returning dummy output so the run continues rather than collapsing. The weights are published under an MIT licence with no regional restrictions, so this is a model you could run yourself; the copy served here is Z.ai's.