Free guide · Aug 26, 2026
Run GLM-5.3-Flash on your own machine
An honest read of the benchmarks, plus three working ways to run it, including my Claude Code setup.
The short version
The 320B/18B split is the whole design. Only 18 billion parameters fire per token, which is why it can be frontier-adjacent and cost fifteen cents a million. It also means the model is cheap to *serve* but not small to *store*, more on that below.
"Running entirely on Chinese AI chips" is their own sentence, not a rumour. The wider GLM-5 family was trained on Huawei Ascend using MindSpore instead of CUDA, with inference also supported on Moore Threads, Cambricon, Kunlun, MetaX, Enflame and Hygon. Whatever else is true, this is a frontier-class model that did not depend on NVIDIA.
42 trillion tokens went through it in six days while it was free and anonymous on OpenCode, making it the second most-used model there behind one that had 56 days to do it.
What's in the full guide
- 01The benchmarks, read honestly
- 02Way 1: Claude Code, via Z.ai directly (easiest, \~5 minutes)
- 03Way 2: Claude Code, via OpenRouter (works with any model, not just this one)
- 04Where your prompts actually go (read this before you pick a route)
- 05Way 3: everything else
- 06Running it locally: the honest version
- 07"Was it self-improving?" No, and here is the real answer
- 08What I would actually do