GLM-4.6: The 200K Context Step in Z.ai's Open Line

What GLM-4.6 is

GLM-4.6 is Z.ai's open-weight model, released under the MIT licence and self-hostable, and it is the release where the GLM line's context window went from 128K to 200K tokens — Z.ai frame that expansion specifically around handling more complex agentic tasks. The roster records 204.8K tokens, which is 200K written out. It sits between GLM-4.5 and the GLM-5 generation, and is superseded by both GLM-5.2 and GLM-5.3, which are covered separately on this site.

What Z.ai changed from GLM-4.5

Z.ai's model card lists five changes from GLM-4.5: the context window grows from 128K to 200K tokens; coding benchmark scores rise, with better results in Claude Code, Cline, Roo Code and Kilo Code, including front-end page generation; reasoning improves and supports tool use during inference; tool-using and search-based agents perform better and integrate more effectively with agent frameworks; and writing aligns more closely with human preferences in style and readability, including role-play. Z.ai evaluated it on eight public benchmarks covering agents, reasoning and coding, reporting gains over GLM-4.5 and comparing it with DeepSeek-V3.1-Terminus and Claude Sonnet 4.

LMArena

On LMArena's text leaderboard GLM 4.6 has 1424 Elo (rank 128); by category, 1458 Elo in Coding (rank 146), 1418 Elo in Math (rank 134), 1439 Elo in Expert (rank 138) and 1401 Elo in Creative Writing (rank 110).

Licensing across the GLM line

Most GLM models on this site are MIT-licensed and self-hostable. GLM-5V Turbo and the older GLM-4-Plus releases are proprietary, and GLM-5.3 had its weights withheld at launch with a release promised two weeks later. The context window reached 1 million tokens at GLM-5.

Inference settings

GLM-4.5 and GLM-4.6 use the same inference method. For general evaluations Z.ai recommend a sampling temperature of 1.0, and for code evaluations such as LiveCodeBench a top_p of 0.95 and top_k of 40. They publish a separate guide for tool-integrated reasoning and a specific tool-call format for search in thinking mode. GLM-4.6 is available through the Z.ai API platform and at chat.z.ai, and its technical report is shared with GLM-4.5.

On the Z.ai API

Z.ai's API documentation gives GLM-4.6 a 200K context window and a 128K maximum output, with text input and output. Z.ai state that GLM-4.6 surpassed Claude Sonnet 4 and other domestic models in 74 real-world coding tests run inside Claude Code, and that it is over 30% more token-efficient than GLM-4.5. The eight benchmarks in Z.ai's comparison include AIME 25, GPQA, LiveCodeBench v6, HLE and SWE-Bench Verified. Z.ai describe it as performing on par with Claude Sonnet 4 on several leaderboards; these are vendor statements.

The GLM line around it

Z.ai's open text releases on this site run from GLM-4.5, with a 128K context, through GLM-4.6 and GLM-4.7 at 200K, to the GLM-5 generation at 1 million tokens. GLM-4.6 and GLM-4.5 share a technical report and an inference method. On LMArena, GLM-4.5 has 1411 Elo (rank 153) and GLM-5.3 has 1481 Elo (rank 28, max effort). GLM-4.6's weights are published on Hugging Face under the MIT licence, and the model is also available at chat.z.ai.

Official Resources