GLM-4.6

What it is

GLM-4.6 is a large language model released by Zhipu in late September 2025; the weights were posted to Hugging Face on 29 September 2025 ^[2][3]. It is an open release: the model is distributed under the MIT License, and the published checkpoint contains about 357 billion parameters ^[1]. A permissive licence plus downloadable weights is the part of the release that matters most to anyone who wants to run, fine-tune or inspect the model rather than only call it through an API.

The model is a direct successor to GLM-4.5, and Zhipu describes the change largely in terms of capability gains rather than a new architecture.

What changed from GLM-4.5

The most concrete change is the context window, which grew from 128K to 200K tokens ^[1]. A longer window lets a model hold more of a codebase, a document set or a long conversation in view at once, which matters for the tasks GLM-4.6 is pitched at.

According to the model card, the release improves on GLM-4.5 in four broad areas ^[1]:

• Coding. Scores on coding tasks are higher, and real-world coding results improve, including for front-end pages.
• Reasoning with tools. Reasoning is stronger when tool use is folded into inference, meaning the model can call outside functions or services while working through a problem.
• Search agents. The model works better as the driver of a search agent.
• Writing and role-play. Output is described as more natural in general writing and in role-play.

The model card also states that GLM-4.6 improves real-world coding results inside tools such as Claude Code, an agentic coding environment in which the model is given a repository and asked to make changes ^[1]. That framing — the model as the engine inside someone else's coding tool — is where much of the practical interest in this release sits.

Evaluation

GLM-4.6 was evaluated on eight public benchmarks covering agents, reasoning and coding ^[1]. Zhipu said the model was competitive with DeepSeek-V3.1-Terminus and Claude Sonnet 4 ^[1]. Those comparisons come from the model's own documentation rather than from an independent third-party evaluation, so they describe the developer's own reading of the results.

Running it

The checkpoint is stored in BF16 and F32 tensor types ^[1]. Beyond that baseline, the release was prepared for lower-precision execution on accelerator hardware: it supports FP8 and Int4 quantization on Cambricon chips and native FP8 on Moore Threads GPUs ^[2]. Support for reduced precision matters in practice because it lowers the memory and compute needed to serve a model of this size, and support for more than one vendor's accelerators widens the set of machines on which it can be deployed.

Where it sits

GLM-4.6 arrived as a mid-cycle refresh of an existing open model line rather than as a new architecture, and it held that position only briefly: it was succeeded by GLM-4.7 in December 2025 ^[2]. Its significance is therefore mostly as a snapshot of what an openly licensed model of roughly 357 billion parameters could do at that point — longer context, stronger coding inside agentic tools, and support for non-NVIDIA accelerator stacks — and as one step in a sequence of releases rather than an endpoint.