GLM 5.2

AI's flagship foundation model built for long-horizon tasks, delivering a solid 1M-token context that stably sustains long, messy coding-agent trajectories. 1 on SWE-bench Pro) and the highest-ranked open-source model across long-horizon coding benchmarks (FrontierSWE, PostTrainBench, SWE-Marathon).

It introduces flexible thinking effort levels (High and Max) to balance capability against latency and compute. 9x at 1M context, and an improved MTP layer for speculative decoding that raises acceptance length by up to 20%.

Released under a pure MIT open-source license with weights available on HuggingFace and ModelScope, it supports transformers, vLLM, SGLang, xLLM, and ktransformers, with 1M input context, 128K max output, thinking mode, function calling, structured output, context caching, and MCP integration.