Hello Blabladores, a lot of you asked me about GLM-5.2 - some even before it was actually available :-) I hear you. In fact, I had the weights downloaded about 15 minutes after it came out (I had a monitor set up for it). I also wanted it running, and I have been trying hard. Even though I had been the whole time in the International Supercomputing Conference this week, I was still working between the booth time, talks, meeetings and networking. You might have seen GLM popping up in and out of Blablador’s model list. That’s me optimizing it. At some point it was working for the websites, but would produce garbage for agents. Some people got upset. I am sorry, I was just as disappointed and confused. Yesterday, with the help of Blablador itself, I got it running in a manner that I got satisfied with the performance and the responses for agents, on 16 H100 gpus, graciously borrowed by the Helmholtz-WestAI project. Thanks, Fritz and Stefan! The problem is that this is a temporary allocation, and this machine is VERY busy. I do not intend to run it continuously there. Our genius colleague Pavel Mezentsev managed to have it running on A100 gpus, even though the model was made to NOT run on this generation of GPUs, and I am trying to learn from him. I have a couple more A100s available where I can drop this model. So, this email is to explain the current situation. TLDR: GLM-5.2 will be appearing and disappearing, as it is on a busy shared machine. It will appear in a more permanent fashion as soon as I return to Jülich and get it running on older equipment I have available for it. That is all. Thank you all for your patience. Let’s bark!® Alex