Zhipu AI today released GLM-4.6, an open-weights language model under the MIT license. The update extends the context window from 128K to 200K tokens and targets agentic coding tasks. Zhipu says the model performs better in applications such as Claude Code, Cline, Roo Code and Kilo Code, and generates more visually polished front-end pages.

On benchmarks, Zhipu reports gains over GLM-4.5 across eight public evaluations covering agents, reasoning and coding. The company claims GLM-4.6 outperforms DeepSeek-V3.2-Exp, though it still trails Claude Sonnet 4.5 on code generation. In the company’s own real-world tests, it wins 48.6% of head-to-head comparisons against Claude Sonnet 4. The model also supports tool use during inference and shows improved reasoning and writing.

The weights are available on Hugging Face, ModelScope, OpenRouter and Z.ai’s API platform. The release comes as Chinese labs continue to push open-source models that challenge Western proprietary systems.

In developer group chats that night, the debate was familiar: another open model that looks good on paper, but how does it actually perform in real coding sessions? Skeptics noted the 48.6% win rate is still a losing record against Sonnet 4, while others pointed to the permissive license and the 200K context as reasons to test it in their pipelines.