I don’t believe in z.ai benchmarks
retained text
All mentioned below ia my experience which can be different from yours. I’m using a lot of OpenSource and closed source models. As a my daily job and as a hobby. And I’m telling you GLM models are shit. Despite of harness or prompt “engineering” - numbers in benchmarks are straight up lie. You can tell that Qwen3.8-27B is a bit benchmaxxed compared to other bigger models. You can suspect deepseek a v4/4.1 flash is benchmaxxed a bit. But GLM-5.2/5.3 being in a leader board - it‘s fucking joke. They are basically lobotomized and I ALWAYS have results worse than with Qwen3.8-flash, deepseek any model, even super stupid gpt5.6-Luna more often giving much better results. Idk why I have problems with them but it can’t be coincidence that this is only models which are not giving what I expect from benchmark numbers. While every other provider/models I use at least somewhat living up to those numbers.
document fetched 2026-09-24 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1wo2hn0/i_dont_believe_in_zai_benchmarks