GLM 4.7

Z.ai's coding partner model, upgraded in two areas over its predecessor: multi-language coding and terminal-agent performance, and more stable multi-step reasoning and execution. It is a Mixture-of-Experts model of roughly 358B parameters covering core coding, tool use and complex reasoning. It introduces interleaved thinking before every response and tool call, preserved reasoning across turns, and per-turn control to switch thinking off for lightweight requests.

GLM-4.7 is the generation before Z.ai moved to the GLM-5 line, and it sits in this catalog as the cheaper, well-understood option rather than as a flagship. Z.ai pitched it as a coding partner, and the gains it claimed over GLM-4.6 are concentrated where a coding agent actually spends its time: 73.8% on SWE-bench, up 5.8 points; 66.7% on the multilingual variant, up 12.9; and 41% on Terminal-Bench 2.0, up 16.5.

The part of this release that has aged best is the thinking control, which is finer-grained than a simple on-or-off switch. Interleaved thinking means the model reasons before every response and every tool call. Preserved thinking means that in coding-agent sessions it carries its own reasoning blocks forward across turns instead of re-deriving them, which is what stops a long session drifting into inconsistency. Turn-level thinking lets a caller disable reasoning for a lightweight request and turn it back on for a hard one, inside the same conversation.

Away from code Z.ai reports gains in tool use and browsing, at 87.4 on the tau-squared agentic benchmark against GLM-4.6's 75.2 and 52 on BrowseComp against 45.1, plus a substantial mathematics and reasoning jump that it summarises as 42.8% on Humanity's Last Exam with tools, up 12.4 points. It also claims plainly better interface output than its predecessor: cleaner, more modern web pages, and slides with more accurate layout and sizing.

Z.ai's own comparison table is where the scope of the model is honest. On Terminal-Bench 2.0 its 41 sits behind Gemini 3.0 Pro's 54.2 and GPT-5.1 High's 47.6, and on the harder terminal set its 33.3 is behind GPT-5.1 High's 43. It was never presented as the strongest agentic model available, only as the strongest for what it cost, which is the reason it is still worth having on the shelf.

The weights are published on Hugging Face and ModelScope with vLLM and SGLang support, so it can be self-hosted; the copy served here is Z.ai's. Reach for it when the task is routine, the volume is high and the rate matters more than the last few points of capability, and reach for the GLM-5 line when it does not.