I very much like the direct demo, similar to how Groq showcased their offerings initially. Would love if they shared more about their custom hardware and model

I very much like the direct demo, similar to how Groq showcased their offerings initially. Would love if they shared more about their custom hardware and model training, as it stands this is far less information than Cerebras and Groq offered upon launch. Model output seems very similar to GLM-4.6 in my purely subjective and very limited testing. Also disappointed to see the "Thinking" tokens fully hidden. Would greatly appreciate a deep dive into what lead to the choice of training their own model, how ground up their model was trained (or whether this is based on an existing model like Composer 1 with Qwen) and why they decided not to provide their hardware with existing models like Cerebras and Groq.