“Thinking Machines plans to release a smaller version of its Inkling model, at about 25% of the parameter count, which is expected to have very good performance.”