GLM-5.3-Flash
Z.ai (Zhipu AI)ships native FP8 weights
implementation detail · filed under inference & serving
The FP8 precision format, mentioned as a model-weight format without a more specific quantization procedure.
The sources on this page name it without explaining how it works. The line above is the extraction's one-line gloss. See inference quantization for the methods filed beside it.
Also called FP8 (8-bit floating point), FP8 quantization, --quantization fp8, FP8 model, FP8 quantized instruct model, Hy4 preview-FP8.
18 quotes from 18 sources, one card per model, each opening with its strongest.
ships native FP8 weights
the default zai-org/GLM-5.3 checkpoint is now native FP8
Hy4 preview-FP8 model weights
Hy4 preview-FP8 | FP8 quantized instruct model
Start the FP8 model
--quantization fp8
--quantization fp8
ships with native FP8 (block-wise e4m3) weights
A separate Hy3-FP8 checkpoint is also released. FP8 lowers the memory footprint for cheaper serving.
Hy3-FP8 | FP8 quantized instruct model
Released under Apache 2.0 with BF16, FP8, NVFP4, and GGUF weights on Hugging Face
For FP8 model ... vllm serve ...
we adopt FP8 inference to improve generation speed and match the trainer
GLM-5 uses FP8 for rollout inference to reduce per-token latency
KV Cache is quantized to FP8
We recommend using the official FP8 checkpoint Qwen/Qwen3.5-397B-A17B-FP8 for optimal serving efficiency.
The examples below cover FP8 and BF16 checkpoints.
Qwen's official FP8 checkpoint