GLM-5.2 leads open models on coding benchmarks

TLDR An open-weight model now matches the closed leaders at serious coding, and it ships MIT-licensed for anyone to use.

The headline: an open-weight model is now the strongest non-proprietary option for serious coding work, and it ships under a license that lets anyone use it freely.

Zhipu released GLM-5.2, a 744-billion-parameter model built for long-horizon coding rather than quick one-off answers. It reads up to a one-million-token context window, enough to hold a whole codebase in view at once, and across three multi-hour coding tests, FrontierSWE, PostTrainBench, and SWE-Marathon, it placed second only to Claude Opus 4.8.

Two technical notes are worth knowing. The model is sparse, meaning only a fraction of those 744 billion parameters fire for any given token, an approach called mixture of experts. And a tweak called IndexShare, which reuses one indexer across every four sparse attention layers, cuts compute per token by 2.9 times at maximum context. Effort-level controls let you trade latency for capability when you need to.

The part that matters most is not the score. Zhipu shipped GLM-5.2 under an MIT license with no regional restrictions, a rare move for a frontier-grade model from a Chinese lab. On standard tests it scores 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, well ahead of GLM-5.1 and closing much of the gap to the closed leaders. Those numbers come from public benchmarks, which are useful but never the whole story.

For anyone weighing whether to build on open models, this narrows the trade-off: you can get close to frontier coding ability without a usage license tying your hands. New to all this? Start with the Beginner's Guide.

Source: VentureBeat