GLM-4.7

What GLM-4.7 is

GLM-4.7 is a large language model from Zhipu, released in December 2025; its weights were posted to Hugging Face on 22 December 2025^[2][3]. Z.ai's own announcement gives 22 December 2025 as the release date^[4]. It appeared in the same period as Zhipu's GLM-4.6V vision model, which was posted on 7 December 2025 alongside a smaller GLM-4.6V-Flash^[3]. The model is built around multilingual agentic coding and terminal tasks — writing and editing code on a user's behalf, and operating a command line — rather than around single-turn question answering^[1].

That focus explains most of its design choices. A coding agent has to hold a plan across many turns, call tools repeatedly, and keep working after an error, so the model's ability to reason between actions matters as much as its raw knowledge.

Thinking control

GLM-4.7 adds three related mechanisms^[1]. Interleaved Thinking means the model thinks before each response and before each tool call, rather than reasoning once at the start and then executing. Preserved Thinking keeps that reasoning available across the turns of a multi-turn agent session, so a long task does not require rebuilding context from scratch at every step. Turn-level thinking control lets the caller decide how much thinking happens at the level of an individual turn, which matters when a session mixes cheap routine steps with hard ones.

Benchmark results

Zhipu reports SWE-bench Verified at 73.8%, a gain of 5.8 points over GLM-4.6, and SWE-bench Multilingual at 66.7%, up 12.9 points^[1]. On Terminal Bench 2.0 it scores 41.0%, up 16.5 points, and on HLE with tools 42.8%, up 12.4 points^[1]. Other reported scores are LiveCodeBench v6 at 84.9, τ²-Bench at 87.4, AIME 2025 at 95.7, GPQA-Diamond at 85.7 and BrowseComp at 52.0^[1]. The largest relative jumps are on the multilingual software-engineering and terminal benchmarks, which is consistent with the stated emphasis on agentic coding across languages.

Z.ai's own evaluation

Beyond standard benchmarks, Z.ai says it evaluated GLM-4.7 on 100 real programming tasks covering frontend, backend and instruction-following, run in a development environment based on Claude Code^[4]. The company also reports placement on two comparison sites: sixth overall in WebDev, where it says GLM-4.7 is the top open model, and first among open-source models on Code Arena^[4]. Z.ai further states that GLM-4.7 performs at or above Claude Sonnet 4.5 on SWE-bench Verified, LiveCodeBench v6 and Terminal Bench 2.0^[4]. These are the vendor's own comparisons and rankings, not independent audits.

License and lineage

GLM-4.7 is released under the MIT License, one of the most permissive standard open-source licenses, which allows reuse and modification with few conditions^[1]. A smaller GLM-4.7-Flash followed in January 2026^[3], and the line was succeeded by GLM-5 in February 2026^[2].

Why it mattered

GLM-4.7 arrived at a point when the practical test for a coding model had shifted from answering programming questions to completing whole tasks in a terminal and a repository, often over many steps. Its contribution was less a new architecture than a set of behaviours tuned for that setting — reasoning interleaved with tool calls, preserved across turns, and controllable per turn — packaged with open weights under a permissive license^[1]. For teams building their own coding agents, that combination is what made the model usable as a component rather than only as a service.