GLM-5.2: 743B MoE, MIT licence, 1M context.

the biggest open-weights release in months. I did the arithmetic on whether you can run it on a 5090. you can't.

first, credit:

GLM-5.2 uses MLA (compressed-latent KV) + DeepSeek sparse attention. that makes its KV cache about 57x https://t.co/as449iiZXx