We’re not far off from the point where a 30B parameter model could do that and run on not-too-expensive hardware. See recent Qwen releases for example and extra

We’re not far off from the point where a 30B parameter model could do that and run on not-too-expensive hardware. See recent Qwen releases for example and extrapolate the current rate of progress from there.