GLM-5.3-FP4

Overview. GLM-5.3-FP4 is GLM-5.3, Z.ai's open-weight mixture-of-experts model, quantized to 4-bit NVFP4. The quantized weights need much less GPU memory than the full-precision model, so it can be served at a lower price per token. It is listed separately from GLM-5.3 so that a request for the full-precision model never lands on the quantized one.

Key strengths. It keeps GLM-5.3's tool use, structured output and long context at a lower price than full precision, with a short time to first token.

When to use it. Use it for chat, drafting, summarisation and agent workflows where cost matters more than the last points of accuracy. For demanding code generation or long chains of reasoning, test it against full-precision GLM-5.3 first.

Limitations. 4-bit quantization can lower output quality, most often on code and long reasoning. Stav has not measured how large the effect is. Benchmark scores shown for this model come from full-precision GLM-5.3 and may overstate this version. Output speed is lower than on full-precision GLM-5.3 endpoints.

Sovereignty note. GLM-5.3 is an open-weight model from Z.ai (China). The weights run on the serving provider's own hardware, and no traffic goes to Z.ai. Where a request is processed depends on the endpoint that serves it.