GLM-4.6

GLM-4.6, published at the end of September 2025, is the release where Z.ai's line became usable for long agent work. The headline change is plain: the context window grew from 128,000 to 200,000 tokens, the size the company has kept for every model since, up to and including the fifth generation. The rest of the announcement covers real-world coding, reasoning with tool use during inference, search agents and writing that reads less like a template. It matters for the catalogue for a second reason: this is the baseline. When Z.ai reports that GLM-4.7 improved SWE-bench Verified by 5.8 points or Terminal Bench by 16.5, GLM-4.6 is the model being improved on. It is also the version that many coding tools standardised on during 2025, which is why it remains in the current price list at $0.60 per million input tokens and $2.20 per million output — the same tier as its successor. The weights are on Hugging Face under an MIT licence. The safetensors index totals 356.8 billion parameters, a mixture-of-experts network of 92 layers with 160 routed experts plus one shared and eight active per token — structurally the same skeleton that GLM-4.7 would reuse three months later, which is worth knowing when comparing the two: the difference between them is training, not size.

▸Videos

No videos yet.

▸Reviews