GLM-4.6 was Zhipu's late-2025 coding workhorse: 357B total parameters, 32B active, a 200K context window, and MIT weights. It earned its reputation inside coding agents — Claude Code, Cline, Roo Code integrations — where it delivered near-frontier coding at a fraction of proprietary cost, using ~15% fewer tokens than GLM-4.5 on comparable tasks.
The GLM-5 family has since taken the flagship slot, but GLM-4.6 remains widely hosted, cheap, and one of the most practical self-host targets among serious coding models — a 4-bit build fits on a 4×80GB node.
GLM-4.6 specifications
| Vendor | Zhipu AI (Z.ai) (China) |
|---|---|
| Architecture | Mixture-of-Experts |
| Total parameters | 357B |
| Active parameters | 32B |
| Context window | 200K tokens |
| Max output | 128K tokens |
| Modalities | text |
| License | MIT — weights free to download, self-host, fine-tune, and use commercially |
| Release date | 2025-10-30 |
| Weights | Hugging Face · GitHub |
Benchmarks and reported results
| Benchmark | Result | Note |
|---|---|---|
| Token efficiency vs GLM-4.5 | ~15% fewer tokens | on comparable agentic tasks |
Running GLM-4.6 locally
357B total / 32B active: 4-bit quantizations run around 200GB — 4×80GB GPUs or a 192GB+ unified-memory machine with offloading. One of the more accessible frontier coding models to self-host.
What GLM-4.6 is best for
- Cost-efficient coding-agent backends
- Self-hosting a serious coding model without trillion-parameter infrastructure
- Teams on the GLM Coding Plan ecosystem
Frequently asked questions
Is GLM-4.6 still worth using after GLM-5.2?
For budget and self-host use cases, yes. GLM-4.6 is smaller (357B vs 753B), cheaper to serve, widely supported, and still strong for coding agents. GLM-5.2 wins when you need the 1M-token context or peak quality.
What hardware does GLM-4.6 need locally?
A 4-bit quantization runs in roughly 200GB — realistic on 4×80GB GPUs (A100/H100) or high-memory Apple Silicon with offloading. Smaller GLM-4.5-Air-class variants serve lighter setups.