GLM 5.2 FP8 with FP8 KV achieves 79.8% on Terminal-Bench 2.1

A user tested the GLM 5.2 model using FP8 precision and FP8 key-value cache quantization on an H200 GPU via SGLang. The evaluation was conducted on the Terminal-Bench 2.1 benchmark suite.

The model passed 71 out of 89 tasks, resulting in a score of 79.8%.
One task timed out and was not rerun, which may have affected the final score.
The test utilized 218,656,815 input tokens with a cache hit rate of 98.8%.

This demonstrates the performance of GLM 5.2 under specific quantization conditions on high-end hardware.

Benchmarks

Benchmark	Model	Score
Terminal-Bench	GLM 5.2 FP8 with FP8 KV	79.8%

Benchmark

Model

Score

Terminal-Bench

GLM 5.2 FP8 with FP8 KV

79.8%