← All models

GLM-4.7 FP8

Zhipu AI (Z.ai) · 2025-12-22 · 358.3B parameters

Z.ai's separately released FP8 checkpoint of GLM-4.7, published with the BF16 weights on December 22, 2025 and used in the official model card's vLLM and SGLang serving examples under the served name glm-4.7-fp8. It stores linear-layer weights as per-channel FP8 (E4M3) via compressed-tensors with dynamic per-token FP8 activations; excluding 120,600,576 quantization-scale elements and 14,400 routing-bias buffer elements, it contains 358,337,776,896 parameters, the same count as the BF16 release. Like GLM-4.7, it is a 92-layer MoE text model with interleaved, preserved, and turn-level thinking controls, a 202,752-position configuration, and an MIT license.

Benchmark scores