GLM 5.1 FP8

GLM-5.1-FP8 is the FP8-quantized variant of GLM-5.1, built for more efficient deployment while preserving the model's 202,752-token context window and 128K output capacity. The FP8 quantization reduces the memory footprint of the full-precision weights, letting teams self-host on systems with less accelerator memory. This variant retains the core capabilities of GLM-5.1, including agentic reasoning, function calling, thinking modes, structured output, and long-horizon task execution, while optimizing for inference speed and memory use. It is compatible with common inference frameworks including vLLM, SGLang, xLLM, Transformers, and KTransformers. Released under the MIT license alongside the full-precision weights, the FP8 build is aimed at teams running GLM-5.1 at scale who want lower memory cost without giving up the coding and agentic performance that defines the family. Context and output limits match the full-precision model (202K input, 128K output), making it a practical choice for resource-constrained, large-scale deployments.