GLM 4.7

GLM-4.7-FP8 is an FP8-quantized build of GLM-4.7 Flash, the 30-billion-parameter MoE coding model, packaged for memory-efficient inference. Quantization reduces the model from roughly 62GB in BF16 to about 30GB in FP8 while keeping sensitive components such as embeddings, layer norms, and router weights in higher precision for near-lossless quality. The variant keeps the coding, agentic, and reasoning behavior of GLM-4.7 Flash and is intended for deployment on FP8-capable GPUs, where it can deliver materially higher throughput than the full-precision weights. It serves teams that want the Flash model's capabilities at lower memory cost for local coding assistants and agentic workflows.